Environmental Factors Affecting the Occurence of Periglacial

Environmental factors aﬀ ecting the occurrence of periglacial landforms in Finnish Lapland: a numerical approach

Jan Hjort

Department of Geography Faculty of Science University of Helsinki

Academic dissertation

To be presented, with the permission of the Faculty of Science of the University of Helsinki, for public criticism in the Auditorium XII of the Main Building (Unioninkatu 34) on May 6th, 2006, at 10 a.m.

I Supervisors

Professor Matti Seppälä Department of Geography University of Helsinki Finland

and

Dr Miska Luoto Finnish Environment Institute, Helsinki / Th ule Institute University of Oulu Finland

Pre-Examiners

Professor Bernd Etzelmüller Department of Physical Geography University of Oslo Norway

and

Professor Charles Harris School of Earth, Ocean and Planetary Sciences Cardiﬀ University United Kingdom

Oﬃ cial Opponent

Reader Julian Murton Department of Geography University of Sussex United Kingdom

Shaker Verlag GmbH, Aachen

II Contents

Abstract VII Acknowledgements VIII List of Figures IX List of Tables XII List of Appendices XIII Abbreviations XIII Symbols XIV

1 INTRODUCTION 1 2 PERIGLACIAL PHENOMENA 5 2.1 Classifi cation of periglacial landforms 6 2.2 Description of periglacial landforms 9 2.2.1 Permafrost landforms 9 2.2.2 Thermokarst features 10 2.2.3 Patterned ground 10 2.2.4 Solifl uction and other slope phenomena 16 2.2.5 Periglacial weathering features 19 2.2.6 Nival phenomena 20 2.2.7 Aeolian processes and landforms 21 3 STUDY AREA 23 3.1 Location and topography 23 3.2 Bedrock and general geology 23 3.3 Weichselian glaciation, deglaciation and soil types 23 3.4 Geomorphology of the Báišduattar – Áilegas 25 3.5 Previous periglacial research in the study region 29 3.6 Past and present climate 30 3.7 Hydrology 32 3.8 Vegetation 32 4 MATERIALS AND METHODS 35 4.1 Modelling data 35 4.1.1 Resolution 35 4.1.2 Periglacial landforms 35 4.1.3 Predictor variables 37 4.1.4 Predictor variable selection and data split 44 4.2 Statistical modelling 44 4.2.1 Statistical formulation 44 4.2.2 Model calibration 46 4.2.3 Model evaluation 46 5 RESULTS 49 5.1 Predictor variables 49 5.2 Periglacial landforms in Báišduattar – Áilegas 49 5.2.1 Palsas 51 5.2.2 Convex non-sorted circles 53 5.2.3 Stony earth circles 53 5.2.4 Earth hummocks 56 5.2.5 Peat pounus 59 5.2.6 Stone pits 60 5.2.7 Sorted nets 61 5.2.8 Sorted stripes 62 5.2.9 Non-sorted solifl uction terraces 64 5.2.10 Sorted solifl uction sheets 64 5.2.11 Sorted solifl uction streams 66 5.2.12 Defl ations 66

V 5.3 Distribution and abundance models 67 5.3.1 Palsas 67 5.3.2 Convex non-sorted circles 71 5.3.3 Stony earth circles 74 5.3.4 Earth hummocks 75 5.3.5 Peat pounus 80 5.3.6 Stone pits 84 5.3.7 Sorted nets 86 5.3.8 Sorted stripes 88 5.3.9 Non-sorted solifl uction terraces 90 5.3.10 Sorted solifl uction sheets 91 5.3.11 Sorted solifl uction streams 94 5.3.12 Defl ations 95 5.3.13 Modelling results: a summary 97 6 DISCUSSION 101 6.1 Periglacial landforms: prevalence, distribution, activity and morphology 101 6.2 Environmental factors affecting periglacial landform occurrence 106 6.2.1 Palsas 106 6.2.2 Convex non-sorted circles 108 6.2.3 Stony earth circles 109 6.2.4 Earth hummocks 110 6.2.5 Peat pounus 111 6.2.6 Stone pits 112 6.2.7 Sorted nets 113 6.2.8 Sorted stripes 114 6.2.9 Non-sorted solifl uction terraces 115 6.2.10 Sorted solifl uction sheets 116 6.2.11 Sorted solifl uction streams 117 6.2.12 Defl ations 118 6.3 Data and methodological issues – advantages and shortcomings 119 6.3.1 Periglacial landform data 119 6.3.2 Predictor data 120 6.3.3 Statistical modelling 122 7 SUMMARY 125 8 CONCLUSIONS 129 REFERENCES 131 OTHER SOURCE MATERIAL 149 APPENDICES 151

VI ABSTRACT

The conclusions about the determinants of earth surface processes and landform patterns are often derived from traditional fi eld survey methods. Recent developments in the spatial and numerical analysing techniques have improved the possibility to study different aspects of geomorphological phenomena in extensive regions. The objective of this research was to map and quantitatively analyse the occurrence of cryogenic phenomena in subarctic Finland in the zone of discontinuous permafrost. More precisely, utilising a grid-based approach the distribution and abundance of periglacial landforms were modelled to identify important landscape scale environmental factors and potential methodological limitations. The study was performed using a comprehensive empirical data set of periglacial landforms from an area of 600 km2 at a 25-ha resolution. The utilised statistical methods were generalized linear modelling (GLM) and hierarchical partitioning (HP). GLMs were used to produce distribution and abundance models and HP to reveal independently the most likely causal variables. The GLM models were assessed utilising statistical evaluation measures, prediction maps, fi eld observations and the results of HP analyses. A total of 40 different landform types and subtypes were identifi ed. At lower altitudes with gentle slope angles occurred earth hummock, stone pit, peat pounu and palsa continuums and at higher altitudes with steeper slopes sorted stripe, solifl uction stream and solifl uction sheet sequences were prevalent. At present, the environmental conditions promote the formation of different cryoturbation and peat accumulation based non-sorted features, whereas most of the sorted landforms were probably formed before the climatic optimum over 8000 years ago. Topographical, soil property and vegetation variables were the primary correlates for the occurrence and cover of active periglacial landforms on the landscape scale. From the pure topographical factors, mean slope angle and mean altitude were commonly in the fi nal models. Peat cover was the most important soil type variable because of its varying thermal properties and moisture holding capacity. Topographical wetness index was a crucial surrogate of environmental factor exhibiting the general soil moisture distribution. From vegetation variables, the shrub cover affected the distribution of several periglacial landforms. In the model evaluation, most of the GLMs were shown to be rather robust although the explanation power, prediction ability as well as the selected explanatory variables varied between the models. The most robust distribution models were constructed with palsa, earth hummock, peat pounu, sorted solifl uction sheet and sorted solifl uction stream data. Earth hummock and peat pounu models obtained the best prediction and explanation ability in the abundance modelling, respectively. The great potential of the combination of a spatial grid system, terrain data and novel statistical techniques to map the occurrence of periglacial landforms was demonstrated in this study. GLM proved to be a useful modelling framework for testing the shapes of the response functions and signifi cances of the variables describing environmental gradients and the HP method helped to make better deductions of the important factors of earth surface processes. Hence, the numerical approach presented in this study can be a useful addition to the current range of techniques available to researchers to map and monitor different geographical phenomena. However, the data related limitations and method-based weaknesses may bias the modelling results and the model outcomes should not be interpreted uncritically.

Keywords: periglacial geomorphology, patterned ground, soliﬂ uction, numerical analyses, generalized linear modelling, logistic regression, hierarchical partitioning, GIS, subarctic, Lapland, Finland

Jan Hjort, Department of Geography P.O. Box 64, FIN-00014 University of Helsinki, Finland

VII ACKNOWLEDGEMENTS

I am grateful to my supervisors Professor Matti Seppälä (Department of Geography, University of Helsinki) and Dr Miska Luoto (Finnish Environment Institute, Helsinki; Thule Institute, University of Oulu) for providing me with the idea and opportunity to do this research. They also have always found time for discussion, advice, comments and constructive criticism throughout the study. In addition, Professors Bernd Etzelmüller and Charles Harris gave several valuable comments on the manuscript. Most of the GIS and statistical analyses as well as the writing were carried out at the Department of Geog- raphy, University of Helsinki, and I wish to express my thanks to all my colleagues there, especially IT Adviser Hilkka Ailio (computing issues and preparation of the layout), PhD student Janne Heiskanen (introduction to the biotope database), PhD student Tommi Sirviö (GIS advice) and IT Administrator Tom Blom (computing problems) as well as PhD student Barnaby Clark who greatly helped with correction of the English text. Kevo Subarctic Research Institute provided good facilities during the fi eld surveys, especially Saini Heino and Kaisu Vierma-Laine who kindly put me up after hard fi eld trips regardless of the day of the week or time. Furthermore, I want to thank Mikko Lantz for fi eld work companion during the summers of 2002 and 2003. I would like to thank the following persons from the Scott Polar Research Institute (SPRI), University of Cam- bridge: Archivist and Curator Robert Headland, Information Assistant Shirley Sawtell and Librarian Heather Lane, who all greatly helped me during the visit to SPRI in May 2005. I wish to express heartfelt special thanks to my wife, PhD student Paula Kuusisto-Hjort, who provided fi eld work companion during the summers of 2002 and 2003, commented on the manuscript at different stages and supported me with love and warmth at every step from the beginning of the study to the fi nishing of the manuscript. The Finnish Geography Graduate school-program has been the main fi nancier of the study. I have also received funding from Seth Sohlberg’s Delegation, Nordenskiöld-samfundet and the Chancellor, University of Helsinki. Finally, I would like to thank all my friends and relatives who have supported me throughout this research that will hopefully be the fi rst step of a long-lasting scientifi c journey.

VIII List of Figures

Figure 1. Simplified conceptual model of the relationships between environmental factors, periglacial processes and landforms commonly found from subarctic regions. 2 Figure 2. Complex palsa mire in Biesjeaggi mire. 10 Figure 3. An example of thawing palsa in Biesjeaggi mire. 11 Figure 4. Small vegetated thermokarst pond in Biesjeaggi mire. 11 Figure 5. Patterned ground, sorted polygon, in a temporary ponding depression. 12 Figure 6. Convex non-sorted circles from the valley northeast of Guivi fell. 13 Figure 7. Sorted circle, debris island, in a blocky soil material in Čeavvresjohka valley. 13 Figure 8. Boulder depression above the treeline in the valley southwest of Guivi fell. 14 Figure 9. Non-sorted steps on a valley slope northwest of Suophášoaivi fell. 15 Figure 10. Non-sorted stripes on a fell slope southeast of Uhc-Áhkováráš fell. 15 Figure 11. Sorted solifluction sheets on the eastern and southeastern slopes of Suobbatoaivi fell. 16 Figure 12. Ploughing block with downslope ridge west of Guivi fell. 17 Figure 13. Large block preventing the soil creep on the slope of Guivi fell. 18 Figure 14. Talus features under rock faces in the Geavvu canyon. 19 Figure 15. Tor on summit close to Suobbatoaivi fell. 20 Figure 16. Large longitudinal nivation hollow-like snow accumulation site north of Čeavrresgielas fell. 21 Figure 17. Deflation depression formed in a sand dune on the eastern slope of Geatgielas fell. 22 Figure 18. Location of the study area in northern Fennoscandia. 24 Figure 19. Generalised map of the study area. 26 Figure 20. Three-dimensional terrain view of the study area. 27 Figure 21. Lateral glacial melt-water channels on the northwestern slope of Gaskkamušaláš fell. 28 Figure 22. Luopmošjáguotkku esker between the Stuorrajávri and Nuorttatjávri lakes in the southern part of the study area. 28 Figure 23. Áhkojohka river valley after the spring flood. 33 Figure 24. Only partly recovered open mountain birch forest damaged by larvae of the moth Epirrita autumnata in the southeastern part of the study area. 34 Figure 25. The main steps of the compilation of response, i.e. periglacial, data. 35 Figure 26. Examples of active and inactive periglacial features. 37 Figure 27. The main interactions between environmental factors affecting periglacial phenomena. 38 Figure 28. Examples of the compiled environmental predictors at 20 m resolution. 40 Figure 29. Examples of the compiled environmental predictors at 25-ha resolution. 41 Figure 30. Generalised geomorphological map of the periglacial phenomena observed in the study area. 53 Figure 31. Number of different periglacial landform types in the study area at 25-ha resolution. 54 Figure 32. String-form palsa in Luopmošjohjeaggi mire. 56 Figure 33. Dome-shaped palsas in Vanađanjeaggi mire. 56 Figure 34. Mean altitude and mean slope angle of the present and absent palsa, convex non-sorted circle and stony earth circle squares. 57 Figure 35. Convex non-sorted circles on a gentle slope west of Ráššoaivi fell. 58 Figure 36. Convex non-sorted circle with small partly vegetated non-sorted polygons on the surface. 58

IX Figure 37. Typical stony earth circle. 59 Figure 38. Active earth hummocks. 59 Figure 39. Mean altitude and mean slope angle of the present and absent earth hummock, peat pounu and stone pit squares. 60 Figure 40. Peat pounus in the western part of Čulloveijeaggi mire. 60 Figure 41. Active stone pit. 61 Figure 42. Active sorted net in a temporary ponding depression. 62 Figure 43. Inactive sorted net from the summit of Ráššoaivi fell. 62 Figure 44. Mean altitude and mean slope angle of the present and absent sorted net, sorted stripe and non-sorted solifluction terrace squares. 63 Figure 45. Active sorted stripes on the slope of Njávgoaivi fell. 63 Figure 46. Small active sorted stripe on valley slope west of Guivi fell. 64 Figure 47. Non-sorted solifluction terrace on the foot of 28° steep slope. 65 Figure 48. Active sorted solifluction sheet with rather angular material and frontal soil embankment. 65 Figure 49. Mean altitude and mean slope angle of the present and absent sorted solifluction sheet, sorted solifluction stream and deflation squares. 66 Figure 50. Active sorted solifluction streams on the fell slope north of Guivi fell. 67 Figure 51. Deflation surface above the treeline in the northwestern part of the study area. 67 Figure 52. Results of the univariate analyses for palsa distribution. 68 Figure 53. Relationship between palsas and environmental variables. 69 Figure 54. Observed and predicted distribution of palsas in the whole modelling area. 70 Figure 55. Results of the hierarchical partitioning analyses for palsa distribution. 70 Figure 56. Results of the univariate analyses for palsa abundance. 71 Figure 57. Evaluation of the final calibration model for the palsa abundance with regression residuals. 71 Figure 58. Results of the hierarchical partitioning analyses for palsa abundance. 71 Figure 59. Results of the univariate analyses for convex non-sorted circle distribution. 72 Figure 60. Relationship between the convex non-sorted circle occurrence and mean altitude. 73 Figure 61. Observed and predicted distribution of convex non-sorted cirlces in the whole modelling area. 73 Figure 62. Results of the hierarchical partitioning analyses for convex non-sorted circle distribution. 73 Figure 63. Results of the univariate analyses for convex non-sorted circle abundance. 74 Figure 64. Evaluation of the final calibration model for the convex non-sorted circle abundance with regression residuals. 75 Figure 65. Results of the hierarchical partitioning analyses for convex non-sorted circle abundance. 75 Figure 66. Results of the univariate analyses for stony earth circle distribution. 75 Figure 67. Relationship between stony earth circles and mean altitude. 76 Figure 68. Observed and predicted distribution of stony earth circle in the whole modelling area. 77 Figure 69. Results of the hierarchical partitioning analyses for stony earth circle distribution. 77 Figure 70. Results of the univariate analyses for earth hummock distribution. 78 Figure 71. Relationship between earth hummocks and environmental variables. 78 Figure 72. Observed and predicted distribution of earth hummock in the whole modelling area. 79 Figure 73. Results of the hierarchical partitioning analyses for earth hummock distribution. 79 Figure 74. Results of the univariate analyses for earth hummock abundance. 79 Figure 75. Results of the hierarchical partitioning analyses for earth hummock abundance. 80 Figure 76. Results of the univariate analyses for peat pounu distribution. 80 Figure 77. Relationship between peat pounu distribution and environmental variables. 81

X Figure 78. Observed and predicted distribution of peat pounus in the whole modelling area. 82 Figure 79. Results of the hierarchical partitioning analyses for peat pounu distribution. 82 Figure 80. Results of the univariate analyses for peat pounu abundance. 83 Figure 81. Relationship between peat pounu abundance and mean slope angle. 83 Figure 82. Results of the hierarchical partitioning analyses for peat pounu abundance. 83 Figure 83. Results of the univariate analyses for stone pit distribution. 84 Figure 84. Relationship between stone pits and glacigenic deposit cover and mean slope angle. 84 Figure 85. Observed and predicted distribution of stone pits in the whole modelling area. 85 Figure 86. Results of the hierarchical partitioning analyses for stone pits distribution. 86 Figure 87. Results of the univariate analyses for sorted net distribution. 86 Figure 88. Relationship between sorted nets and peat cover. 87 Figure 89. Observed and predicted distribution of sorted nets in the whole modelling area. 87 Figure 90. Results of the hierarchical partitioning analyses for sorted net distribution. 88 Figure 91. Results of the univariate analyses for sorted stripe distribution. 88 Figure 92. Relationship between sorted stripes and mean altitude and mean slope angle. 89 Figure 93. Observed and predicted distribution of sorted stripes in the whole modelling area. 89 Figure 94. Results of the hierarchical partitioning analyses for sorted stripe distribution. 90 Figure 95. Results of the univariate analyses for non-sorted solifluction terrace distribution. 90 Figure 96. Observed and predicted distribution of non-sorted solifluction terraces in the whole modelling area. 91 Figure 97. Results of the hierarchical partitioning analyses for non-sorted solifluction terrace distribution. 92 Figure 98. Results of the univariate analyses for sorted solifluction sheet distribution. 92 Figure 99. Relationship between sorted solifluction sheets and environmental variables. 92 Figure 100. Observed and predicted distribution of sorted solifluction sheet in the whole modelling area. 93 Figure 101. Results of the hierarchical partitioning analyses for sorted solifluction sheet distribution. 94 Figure 102. Results of the univariate analyses for sorted solifluction sheet abundance. 94 Figure 103. Results of the hierarchical partitioning analyses for sorted solifluction sheet abundance. 94 Figure 104. Results of the univariate analyses for sorted solifluction stream distribution. 95 Figure 105. Relationship between the sorted solifluction stream occurrence and mean altitude and mean slope angle. 95 Figure 106. Observed and predicted distribution of sorted solifluction streams in the whole modelling area. 96 Figure 107. Results of the hierarchical partitioning analyses for sorted solifluction streams distribution. 97 Figure 108. Results of the univariate analyses for deflation distribution. 97 Figure 109. Relationship between deflations and sandy soils and altitude. 98 Figure 110. Observed and predicted distribution of deflations in the whole modelling area. 98 Figure 111. Results of the hierarchical partitioning analyses for deflation distribution. 99 Figure 112. Summary of the environmental factors of the final distribution models. 100 Figure 113. Summary of the environmental factors of the final abundance models. 100 Figure 114. Schematic cross-section summarising the study results. 102 Figure 115. Potentially important environmental factors affecting periglacial feature distribution in northernmost Finnish Lapland at a landscape scale. 127 Figure 116. Potentially important environmental factors affecting periglacial feature abundance in northernmost Finnish Lapland at a landscape scale. 128

XI List of Tables

Table 1. Characteristic geomorphic processes in periglacial environments and examples of the resulting landforms. 5 Table 2. Washburn’s classification system of periglacial phenomena. 7 Table 3. Åkerman’s classification system of periglacial phenomena. 7 Table 4. Ballantyne and Harris’ classification system of periglacial phenomena. 8 Table 5. French’s classification system of periglacial phenomena. 9 Table 6. Summary of the climate parameters measured at the Kevo Meteorological Station. 31 Table 7. Pure topographical variables and their description. 39 Table 8. Surrogates for soil moisture and their description. 42 Table 9. Temperature and solar radiation variables and their description. 42 Table 10. Variables, coefficients, standard errors, t- and p-values of the final model for minimum air temperature. 42 Table 11. Soil type variables and their description. 43 Table 12. Vegetation variables and their description. 43 Table 13. Spatial variables and their description. 44 Table 14. Correlation matrix of the environmental predictors used in the statistical analyses. 50 Table 15. Periglacial landforms and their prevalence in the Báišduattar – Áilegas area at the 25-ha modelling resolution. 55 Table 16. Topographical characteristics of the active landform occurrences in relation to absent sites. 57 Table 17. Variables, coefficients, standard errors, z- and p-values of the final model for palsa distribution. 68 Table 18. Deviance information and degrees of freedom of the final distribution and abundance models. 69 Table 19. Variables, coefficients, standard errors, t- and p-values of the final model for palsa abundance. 71 Table 20. Variables, coefficients, standard errors, z- and p-values of the final model for convex non-sorted circle distribution. 72 Table 21. Variables, coefficients, standard errors, t- and p-values of the final model for convex non-sorted circle abundance. 74 Table 22. Variables, coefficients, standard errors, z- and p-values of the final model for stony earth circle distribution. 76 Table 23. Variables, coefficients, standard errors, z- and p-values of the final model for earth hummock distribution. 78 Table 24. Variables, coefficients, standard errors, t- and p-values of the final model for earth hummock abundance. 80 Table 25. Variables, coefficients, standard errors, z- and p-values of the final model for peat pounu distribution. 81 Table 26. Variables, coefficients, standard errors, t- and p-values of the final model for peat pounu abundance. 83 Table 27. Variables, coefficients, standard errors, z- and p-values of the final model for stone pit distribution. 85 Table 28. Variables, coefficients, standard errors, z- and p-values of the final model for sorted net distribution. 86 Table 29. Variables, coefficients, standard errors, z- and p-values of the final model

XII for sorted stripe distribution. 88 Table 30. Variables, coefficients, standard errors, z- and p-values of the final model for non-sorted solifluction terrace distribution. 90 Table 31. Variables, coefficients, standard errors, z- and p-values of the final model for sorted solifluction sheet distribution. 92 Table 32. Variables, coefficients, standard errors, t- and p-values of the final model for sorted solifluction sheet abundance. 94 Table 33. Variables, coefficients, standard errors, z- and p-values of the final model for sorted solifluction stream distribution. 96 Table 34. Variables, coefficients, standard errors, z- and p-values of the final model for deflation distribution. 97 Table 35. Summary of the model performances of the distribution models. 99 Table 36. Summary of the model performances of the abundance models. 99

List of Appendices

Appendix 1. Distribution of palsas in the study area. 151 Appendix 2. Distribution of convex non-sorted circles in the study area. 152 Appendix 3. Distribution of stony earth circles in the study area. 153 Appendix 4. Distribution of earth hummocks in the study area. 154 Appendix 5. Distribution of peat pounus in the study area. 155 Appendix 6. Distribution of stone pits in the study area. 156 Appendix 7. Distribution of sorted nets in the study area. 157 Appendix 8. Distribution of sorted stripes in the study area. 158 Appendix 9. Distribution of non-sorted solifluction terraces in the study area. 159 Appendix 10. Distribution of sorted solifluction sheets in the study area. 160 Appendix 11. Distribution of sorted solifluction streams in the study area. 161 Appendix 12. Distribution of deflations in the study area. 162

Abbreviations

AML ARC Macro Language a.s.l. above sea level AUC Area Under the Curve BP before present CTA Classification Tree Analysis DEM Digital Elevation Model D.f. Degrees of freedom GIS Geographical Information System GLM Generalized Linear Model GPS Global Positioning System HP Hierarchical Partitioning LP Linear Predictor LS Least Square MAAT Mean Annual Air Temperature MARS Multiple Adaptive Regression Splines max maximum

XIII min minimum n.s. not significant PCC Percentage of Correct Classification Q–Q Quantile–Quantile RMSE Root-Mean-Square Error RMSPE Root-Mean-Square Prediction Error ROC Receiver Operating Characteristic RS Remote Sensing SE Standard Error ssp. subspecies std standard deviation

Symbols

α slope angle / constant As upslope contributing area β regression coefficient E Elevation–relief ratio ei distance difference between test points exp exponential I independent contribution J cojoint contribution κ Cohen’s Kappa statistic L% percentage of lakes ln natural logarithm log logarithm μ expected value n number of observations ω topographical wetness index pi probability p statistical significance Rs Spearman’s rank correlation coefficient t test value |x| absolute value x Finnish National Grid Coordinate (east) xi predictor variable y Finnish National Grid Coordinate (north) yλ Box-Cox transformation z test value Z elevation * p < 0.05 ** p < 0.01 *** p < 0.001

XIV

1 INTRODUCTION

Determination of the environmental factors graphical information system (GIS) science, controlling earth surface processes and land- remote sensing (RS) and statistical techniques form patterns is one of the central themes have improved the possibility to survey ex- in physical geography (e.g. Goudie 1995; Al- tensive regions and study different aspects of len 1997). Geomorphological processes are earth surface processes in the cold regions (e.g. often studied utilising traditional fi eld survey Walsh et al. 1998; Etzelmüller et al. 2001; Gru- methods (e.g. Verstappen 1983; Summerfi eld ber & Hoelzle 2001; Luoto & Seppälä 2002a; 1991). However, the identifi cation of the main Grosse et al. 2005; Gurney & Bartsch 2005). drivers of the geomorphological processes GIS offers an analytical framework for is often challenging, particularly if complex storing, combining, displaying and analysing systems and extensive areas are under inves- large data sets on multiple spatial scales (e.g. tigation. Novel spatial analysis and modelling Walsh et al. 1998; Burrough & McDonnel methods in combination with thoughtful geo- 2000). Correspondingly, statistical methods morphological understanding could provide provide a mathematical basis for the interpre- new insights into the process-environment tation of relationships between response and relationships (e.g. Atkinson et al. 1998; Bledsoe predictor variables and they enable the explo- & Watson 2001; Lewkowicz & Ednie 2004; ration of roles of different environmental cor- Luoto & Hjort 2005, in press; Ayalew & Yam- relates (Atkinson et al. 1998; Guisan et al. 2002; agishi 2005; cf. Wilcock & Iverson 2003). Luoto et al. 2004a). Furthermore, the numeri- Human activity in the polar and alpine re- cal techniques provide an effi cient approach to gions has increased constantly (Harris 1986; summarise the geomorphological knowledge Davis 2001). Moreover, the increasing atmos- and help to draw general conclusions of the pheric concentrations of greenhouse gases studied phenomena for academic, applied and could lead to signifi cant changes in regional economical purposes. However, several poten- and seasonal climatic patterns. Because of the tial data and method-based shortcomings may major infl uence of climate on the activity of affect the reliability of the results when geo- periglacial processes from the continental to graphical data sets are used in statistical mod- local scale, climate change could strongly in- elling (Clark & Hosking 1986: 17–19; Legen- fl uence the geomorphological processes in dre 1993; MacNally 1996; Luoto & Heikkinen subarctic areas, particularly in the zone of 2003). discontinuous permafrost (Anisimov & Fit- Periglacial phenomena operate over a wide zharris 2001; Nelson et al. 2001; Nelson 2003). range of scales (Washburn 1979; French Observations and experiments also indicate 1996). On a large scale, periglacial processes climate-induced changes in periglacial proc- have been described in relation to latitudinal esses (e.g. Christensen et al. 2004; Lewkowicz and longitudinal environmental gradients, i.e. & Harris 2005). Consequently, the need for climatic conditions (e.g. Lundqvist 1962; Har- detailed distributional information of the geo- ris S.A. 1982a). On a fi ne scale, local topo- morphological phenomena governing the cold graphical, soil and vegetation factors have environments has recently increased consider- been used to determine the distribution and ably. activity of phenomena (e.g. van Vliet-Lanoë Despite that periglacial processes and land- 1988a; Matthews et al. 1998). The phenomena- forms have been intensively studied for over a environment interactions on a landscape scale century (French 2003), gaps in our knowledge are, however, still rather poorly explored in exist (e.g. Clark 1988; Ballantyne & Harris periglacial studies. 1994). The developments in the fi eld of geo- In this study, a quantitative description

1 Introduction and statistical analysis of the distribution and the models and (6) explore the methodologi- abundance of different active periglacial land- cal advantages as well as limitations. Moreo- forms on the landscape scale is provided. This ver, utilising two alternative multivariate tech- work is based on a grid-based approach at the niques, namely generalized linear modelling resolution of 500 x 500 m (25-ha). The general (GLM) and hierarchical partitioning (HP), the aim is to assess the applicability of the numeri- aims are to gain deeper insight into the proc- cal modelling techniques in combination with ess-environment relationships and study reli- GIS data to analyse the main determinants ability of the results when geographical data of periglacial phenomena. More precisely, sets are used in statistical modelling. However, the objectives are to (1) map the occurrence the determination of scale dependency of of periglacial landforms in subarctic Finland the phenomena occurrence and potential ef- (Seppälä 1997a: 83), (2) study statistically the fects of climate change on the processes and association between landforms and environ- landforms are beyond the scope of this study. mental factors driven from GIS data sets (Fig. In general, spatial scaling issues have received 1; Etzelmüller et al. 2001: 89), based on the increasing attention in GIS science during re- previous, (3) determine the most important cent decades but also in geomorphology (e.g. environmental factors affecting the distribu- Luoto & Hjort in press). tion and abundance of periglacial phenomena The study is performed using a compre- on the landscape scale (e.g. Barsch 1993: 154, hensive empirical data set of periglacial land- 159–160), (4) identify the shapes of the re- forms. The focus is on the active features sponses, (5) evaluate the prediction abilities of because many of the utilised predictor vari-

Figure 1. Simpliﬁ ed conceptual model of the relationships between environmental factors, periglacial processes and periglacial landforms commonly found from subarctic regions, particularly northern Fennos- candia (e.g. Lundqvist 1962; Harris C. 1982; Jahn & Siedlecki 1982; Meier 1987; Seppälä 1987, 2005a; for more details see Chapter 2). The presented processes are grouped including several different subtypes of periglacial processes. For example, rapid mass movements are different kind of falls, ﬂ ows, avalanches and slides and cryoturbation refer to the general lateral and vertical displacement of soil due to freeze-thaw activity (e.g. French 1988).

2 Introduction ables illustrate present environmental condi- Speciﬁ c questions to be answered are: (1) tions and thus the current activity (cf. Nies- What kind of periglacial phenomena exist in sen et al. 1992: 188). The subarctic study area the study area? (2) Which are prevalent land- of 600 km2 is chosen based on the previous form types in subarctic Finland? (3) What are knowledge on the existence of numerous per- the most important landscape scale environ- iglacial landforms and variability of the envi- mental factors of the periglacial phenomena? ronmental conditions in the region (e.g. Piirola (4) How well can we predict the distributions 1969, 1972; Kejonen 1979: Luoto & Seppälä and abundances of different periglacial land- 2000; van Vliet-Lanoë & Seppälä 2002). The forms on the basis of GIS data? (5) Are the diversity of relief, soil properties and vegeta- model outcomes realistic? (6) Which are criti- tion provides an excellent setting for detailed cal methodological limitations? (7) What are quantitative study of the determinants of per- the next critical steps to improve the models iglacial landforms and processes in subarctic in periglacial research? landscapes.

2 PERIGLACIAL PHENOMENA

Which are periglacial phenomena? How do frost processes (e.g. Tricart 1968: 830; Jahn we defi ne a periglacial environment? Which 1975: 10–14; Ballantyne & Harris 1994: 28). factors control the activity of geomorpho- However, because there is no agreement for logical processes in cold environments? These the quantitative limits of periglacial environ- and numerous similar questions have inspired ments, it is widely accepted that the term per- geoscientists even long before the concept of iglacial refers to the conditions and phenom- periglacial was considered (e.g. Jahn 1975: 1–4; ena associated with cold, non-glacial regions Ballantyne & Harris 1994: 4–5). The term per- where frost processes dominate (Karte 1979: iglacial was fi rst introduced by von Łoziński 177; Washburn 1979: 4; Ballantyne & Harris (1909) to describe frost-weathering conditions 1994: 3; French 1996: 3; van Everdingen 2005: in the Carpathian Mountains. Ever since re- 54–55). Regardless of the different defi ni- searchers have made an effort to characterise tions, we can emphatically state that periglacial the periglacial domain but we cannot fi nd any conditions prevail in extensive areas and even generally accepted quantitative parameters to 20–35% of the Earth’s land surface could be the defi nition of periglacial phenomena or en- considered as periglacial environment (French vironments. 1996: 5; Worsley 2004: 773). Different, mainly climate-based, limits have The processes acting in the periglacial do- been presented. According to Peltier (1950: main and responsible for the formation of dif- 215) mean annual air temperature (MAAT) ferent landforms are numerous (Fig. 1; Table should be between −15°C and −1°C and 1). Many of the so-called periglacial processes precipitation should range from ca. 130 mm can also operate in the other non-glacial en- to 1400 mm. Wilson (1968: 723) gave values vironments and differ only in their frequency −12°C – +2°C and 50 – 1250 mm for the and/or intensity. Characteristic periglacial phe- limits. French (1996: 20) defi ned a periglacial nomena are related to the permafrost aggrada- region to be all those areas where the MAAT tion and degradation (e.g. French 1996: 6) but, is under +3°C. In addition to pure climato- in general, frost action is the most widespread logical determinants, there exist defi nitions and important periglacial process (Washburn based on permafrost, treeline, snowcover and 1979: 6).

Table 1. Characteristic geomorphic processes in periglacial environments and examples of the resulting landforms (Washburn 1979; French 1996).

Processes Landforms Ground ice development (e.g. ice intrusion and segregation) Pingos Thermal, desiccation and dilation cracking of soil Non-sorted polygons Frost and complex physico-chemical weathering processes Block fi elds Frost heave, frost thrusting, mass displacement and particle size sorting Sorted circles Thawing of ice-rich soil i.e. thermokarst Thermokarst depressions Snowbank related processes (nivation) Cryoplanation terraces Soil creep and gelifl uction (solifl uction) Solifl uction lobes Rapid mass movements (e.g. slushfl ows, debris fl ows and rockfalls) Taluses Fluvial processes (e.g. associated with spring fl oods) Break-up features Aeolian activity caused by strong winds Sand dunes Coastal processes associated with lake or sea ice Ice-shove ridges

5 Periglacial phenomena

The environmental determinants infl uenc- 2.1 Classifi cation of periglacial ing the phenomena are almost as varied as the landforms periglacial processes themselves (e.g. Fig. 1). Washburn (1979: 10–17) listed fi ve basic (i.e. independent) factors and three dependent determinants. In general, the most important en- Classifi cation of the landforms is a fairly chal- vironmental drivers are climate (temperature, lenging task due to the diversity of the perigla- precipitation and wind), topography, material cial phenomena (e.g. Ballantyne & Harris 1994: (surfi cial soil and bedrock), moisture condi- 7–8). The classifi cation can be performed tions, snow- and ice cover as well as vegeta- based on the processes (e.g. frost, solifl uction, tion. Washburn (1979: 15) also included time aeolian), landforms (e.g. patterned ground), and human activity into the set of factors. environments (e.g. forest, mire, slope, low- Climatological conditions affect, for example, land), climate (e.g. Harris S.A. 1982a) or some freeze-thaw cycles and general moisture distri- combination of the former. Most of the re- bution and therefore all periglacial processes. searchers have created a classifi cation of their In addition, the wind has a signifi cant role in own to structure their work (e.g. Åkerman the snow redistribution, which affects the soil 1980). Consequently, there is not available any temperatures and moisture conditions. Soil classifi cation that includes all known landform moisture is among temperature as the most types. This is natural, because every study area important determinant of frost processes is unique and probably no region is so repre- (Fig. 1; Washburn 1979: 63). Soil material af- sentative that it would include all known per- fects the moisture distribution and process iglacial features. activity. Topography, snowcover and vegeta- Six different classifi cations will be briefl y tion infl uence the processes through the main presented here. Five of them are taken from drivers; temperature and soil moisture. How- periglacial textbooks (Embleton & King 1975; ever, topography affects directly on the slope Washburn 1979; Williams & Smith 1989; Bal- phenomena and vegetation on the slope and lantyne & Harris 1994; French 1996) and one aeolian processes. is from a mapping research (Åkerman 1980). Subsequently, different periglacial landform The classifi cations were interpretated from the classifi cations will be presented. In general, the sources because explicit tables describing the classifi cation of diverse and continuous phe- classifi cations were not available. nomena areas where activity may vary spatially Embleton and King (1975) classify perigla- and temporally is rather diffi cult and artifi cial, cial phenomena into six main groups: (1) fro- but the classifi cation of phenomena is essential zen ground phenomena, (2) patterned ground, to structure the study and in the compilation (3) mass movement and slope deposits, (4) the of response data sets. Description of differ- action of snow, (5) cryoplanation, tors, block- ent periglacial landforms is provided after the fi elds and blockstreams and (6) wind action. classifi cation. Focus will be on the landform Their classifi cation is partly process and partly types commonly found in subarctic regions, landform based. Washburn’s (1979) classifi - particularly in Finland (see also Okko 1954; cation is mostly based on the processes but Aartolahti 1980; Seppälä 1997a). Interaction periglacial forms such as patterned ground, between environmental factors and periglacial involutions and palsas are grouped together phenomena are treated in Chapter 6 in more (Table 2). Åkerman’s (1980) classifi cation is detail. also mainly based on the processes but he has more classes than the previous classifi cations (Table 3). Williams and Smith (1989) divided landforms into only two main groups according to the topography of the occurrence area, i.e. landforms that prevail on slopes and sub- sides and features on level ground. Ballantyne

6 Periglacial phenomena

Table 2. Washburn’s (1979) classiﬁ cation system of periglacial phenomena.

Main division Periglacial landforms Periglacial forms Patterned ground, involutions, stone pavements, string bogs, palsas and pingos Mass-wasting processes Avalanche features, slushfl ow deposits, frost creep and gelifl uction deposits, ploughing blocks and braking blocks, rock glaciers, talus and protalus ramparts Nivation Nivation benches and hollows and cryoplanation terraces Slopewash Grèzes litées Fluvial action Icing, break-up features, fl at-fl oored valleys, asymmetric valleys, dry valleys and dells Lacustrine and marine Varves, ice rafting features, lacustrine and marine ice-shove ridges action Wind action Loess, dunes, niveo-eolian forms, defl ation and ventifacts Thermokarst Collapse pingos, thaw slumps, linear and polygon troughs, beaded drainage, thaw lakes and alases

and Harris (1994) use a partly process and and rather incomplete because fl uvial process- partly landform-based classifi cation also in- es were largely lacking and aeolian and coastal cluding the spatial and temporal aspect (Table phenomena were absent. Washburn (1979) 4). French (1996) has six classes and he uses treated periglacial phenomena relatively com- mainly the process-based approach (Table 5). pletely but the classifi cation differed slightly All the presented classifi cations more or from the others. The classifi cations of Emble- less differ from each other. The classifi cation ton and King (1975), Ballantyne and Harris of Williams and Smith (1989) was the simplest (1994) and French (1996) were relatively simi-

Table 3. Åkerman’s (1980) classiﬁ cation system of periglacial phenomena.

Main division Periglacial landforms Weathering processes Block fi elds and slopes, cavernous weathering and surface karst features Ground ice features Palsas, hummocks, ice-wedge polygons, “talus pingoids”, “pseudo pingos”, hydrolaccolites, naledi and naledi pavements Thermokarst processes Thermokarst basins, slope of thermokarst denudations, thermokarst landslides, ravines and dolines Frost processes Circles, nets, polygons, steps and stripes Wind action Dunes (e.g. embryonal and tail dunes), unspecifi ed sandy deposits, ventifacts and wind erosion in vegetation Nivation processes Nivation hollows, snow-patch pavements and perennial snow-patches Slope processes Rock-fall scars, talus cones, debris fl ow tracks and accumulations, protalus ramparts, avalanche tracks and accumulations, gelifl uction sheets and lobes, block streams, ploughing and braking blocks and land slide scars Coastal and littoral Cliff and cliff with caves, stacks, edge of abrasion platforms, seaweed processes accumulations, ice push ridges, thermokarst in buried ice, active beaches (sandy) and driftwood accumulations Fluvial processes Small streams, water falls, fords, rapids, ravines, sink holes, wells and active alluvial deposits

7 Periglacial phenomena

Table 4. Ballantyne and Harris’ (1994) classiﬁ cation system of periglacial phenomena.

Lowland Britain Main division Periglacial landforms Ice-wedges and tundra Ice wedges and ice-wedge casts, tundra polygons, sand wedges and polygons sand-wedge polygons and soil wedges Pingos and related ground- Pingos, pingo ramparts and scars, palsas and mineral palsas, seasonal ice phenomena frost mounds, icing and thermokarst depressions (alases, thaw lakes, ground-ice slumps) Active layer processes: Circles, polygons, irregular networks, stripes, step-like, oval, lobate and cryoturbation and patterned garland patterns and involutions ground Mass wasting and slope Skinfl ows and active-layer detachment failures evolution Mudfl ows and debris fl ows Ground-ice slumps Slopewash (colluvium, gelifl uctates) Fluvial and aeolian processes Asymmetrical and symmetrical valleys Dells and braided river channels Loess, dunes, coversands, ventifacts Niveo-aeolian deposits

Upland Britain Main division Periglacial landforms Frost weathering Block fi elds, slopes and streams, debris-mantled slopes and tors Patterned ground Circles, polygons etc. (see above) Solifl uction landforms Solifl uction lobes, steps, terraces and sheets, boulder sheets and lobes and ploughing boulders Talus slopes and related Talus slopes (sheets and cones), avalanche cones, tongues and tracks, landforms debris cones and fl ows, protalus ramparts and rock glaciers (protalus and morainic rock glaciers) Nival, fl uvial, aeolian and Nivation hollows and benches, cryoplanation erraces, stratifi ed slope coastal features deposits (grèzes litées and éboulis ordonnées), defl ation surfaces and pavements, wind-patterned ground (defl ation scars, wind stripes and crescents), turf-banked and defl ation terraces and periglacial shorelines

lar. They fi rst introduced permafrost and relat- observations of continuous permafrost re- ed phenomena, then active layer processes and gions where the phenomena can differ from patterned ground, thirdly slope processes and the marginal periglacial areas processes [e.g. fi nally fl uvial, aeolian, coastal and weathering two-sided freezing (French 1988: 151)]. How processes and landforms in different orders. proper are these systems then for classifying Embleton and King (1975) had defi ciencies periglacial landforms in different cold envi- in their system because the coastal and fl uvial ronments, for example in the subarctic Fenno- phenomena were not classifi ed, although these scandia? In general, it is probable that any of were discussed briefl y in the introduction. the above presented classifi cations would be Åkerman (1980) treated all common perigal- suitable but every research may need a some- cial phenomena and the distinction to the pre- how adjusted approach because periglacial vious was only in the order of themes. phenomena vary geographically. Most of the classifi cations are based on the The classifi cation utilised in this study was

8 Periglacial phenomena

Table 5. French’s (1996) classifi cation system of periglacial phenomena. Main division Periglacial landforms Permafrost and ground ice Palsas, rock glaciers, seasonal frost mounds, ice wedges, pingos and ice-cored mounds (e.g. Hydrolaccoliths) Thermokarst Alases, ice-wedge thermokarst terrain, retrogressive thaw slumps, thaw lakes and depressions The active layer Patterned ground (circles, polygons, nets, steps and stripes) Hillslope precesses Solifl uction sheets and lobes, ploughing blocks, slopewash (nivation, Grèzes litées), debris fl ows and avalanches Fluvial processes and Icing, braided channels, periglacial sandurs and asymmetrical valleys landforms Wind action and coastal Ventifacts, stone pavements, defl ation depressions, loess, sand dunes processes and sheets, ice-push features, cliffs and cold climate deltas adapted from the systems of Embleton and pingo-like features and rockglaciers can be King (1975), Ballantyne and Harris (1994; Ta- found from neighbouring subarctic regions ble 4) as well as French (1996; Table 5). How- (e.g. Østrem 1971; Åkerman & Malström ever, the differences to Washburn (1979; Ta- 1986; Lagerbäck & Rodhe 1986). Permafrost ble 2) and Åkerman (1980; Table 3) are minor. has also been found in small peaty earth hum- The main division of the landforms is based mocks called pounus (e.g. Ruuhijärvi 1962; on processes, although patterned ground fea- Piirola 1972; Seppälä 1998; Luoto & Seppälä tures are grouped together. The subdivision of 2002b) as well as in bedrock (King & Seppälä the patterned ground group is based on Wash- 1987, 1988; Kukkonen & Safanda 2001) and burn (1956, 1979) with minor alterations (van minerogenic ground (Hirvas et al. 2000, 2005; Everdingen 2005: 53). Moreover, sorted fi elds, Vanhala et al. 2005). such as boulder depressions, are included in Palsas are 0.5–10 m high permanently fro- this class (see Lundqvist 1962). In the group zen peat or partly mineral soil hummocks that of slope features, the terms sorted and non- occur widely in the circumpolar discontinu- sorted are used instead of turf- and stone ous or sporadic permafrost zone with slightly banked to describe the general sorting of the continental climate and with thin snowcover solifl uction landforms, as recommended by (Fig. 2; Lundqvist 1969; Seppälä 1972a, 1988, Ballantyne and Harris (1994: 208–209). 2004a; Moore 1984; Nelson et al. 1992; Gur- ney 2001). Åhman (1977: 38–42) has described fi ve different morphological types of 2.2 Description of periglacial palsas and palsa mires: (1) 1–1.5 m high and landforms extensive palsa plateaus, (2) 2–6 m high and 50–500 m long esker or ridge-form (Seppälä 1988: 255) palsas, (3) 1–2 m high and 25–100 2.2.1 Permafrost landforms m long string palsas, (4) 2–6 m high circular conical or dome-shaped (Seppälä 1988: 255) Permafrost is defi ned as ground (soil or bed- palsas and (5) palsa complexes where palsas of rock) in which the temperature remains at or different morphology and/or in various stages below 0ºC for at least two consecutive years of development exist (Fig. 2). (Muller 1947: 3; van Everdingen 2005: 55). Different theories for palsa formation have Northern Finland is located in the zone of dis- been presented, for example snowcover, vege- continuous and sporadic permafrost (Brown tation and buoyancy hypothesis (for summary et al. 1997) and palsas are the only true per- see Seppälä 1994a; Gurney 2001). The snow- mafrost features (Seppälä 1997b), albeit that cover theory has been experimentally tested

9 Periglacial phenomena

Figure 2. Complex palsa mire with ca. 1.5–2 m high palsas in Biesjeaggi mire (69°27’17”N, 26°8’58”E / 345 m a.s.l. / 15th of July 2003). All photographs were taken by the author in the study area during the summers of 2002–2004.

(Seppälä 1982a; Seppälä 1995a). In general, scale disturbances (French 1996: 109–112). palsas are considered to originate as a result of Thermokarst ponds resulting from the thaw- the growth of segregation ice by cryosuction ing of palsas are frequently found from palsa and their inner structure is mostly composed mires (e.g. Thie 1974; Åhman 1977; Kershaw of segregated ice, frozen peat and silty mineral & Gill 1979; Matthews et al. 1997; Zuidhoff soil with small ice crystals (Seppälä 1988: 263– & Kolstrup 2000; Zuidhoff 2002). In Finland, 266; Fig. 1). The peat cover is essential for the the sizes of the features generally range from survival of the frozen core and the insulation is a few tens of square metres to one hectare and based on the thermoconductivity of peat (see they are water or vegetation covered (Fig. 4; cf. Williams & Smith 1989: Table 4.1, 113). For Luoto & Seppälä 2003). Thawed and collapsed example, Railton and Sparling (1973) and Sep- palsa may leave a circular peat rampart i.e. rim pälä (1979a) have described the cyclic devel- ridge around the thermokarst depression (e.g. opment of the palsas with different aggrada- Svensson 1969; Seppälä 1988: Figure 11.6a). tion and degradation stages (see also Zuidhoff In general, the distribution of the thermokarst & Kolstrup 2005: 57–59). At present, palsas ponds is closely related to the occurrence of are generally degrading throughout their dis- palsas but thermokarst features have wider tribution range (Fig. 3), probably because of vertical and horizontal distribution at the land- the regional climate warming (Matthews et al. scape scale and they can be found from areas 1997; Zuidhoff & Kolstrup 2000; Luoto et al. where palsas have disappeared (Luoto & Sep- 2004b), but recently developed palsas can also pälä 2003). be found (e.g. Seppälä 2003a: 3).

2.2.3 Patterned ground 2.2.2 Thermokarst features Patterned ground landforms are more or less The term thermokarst describes topographic symmetrical structures visible at the ground irregularities, mainly depressions, resulting surface and their size and shape is dependent from the melting of ground ice (Pissart 2004: on the topographical location and formative 1049–1050). Reasons for the permafrost deg- processes (Fig. 5; cf. Troll 1944). No consen- radation are numerous and they can range from sus exists on what features belong to patterned global (e.g. climate change) to local (e.g. forest ground. Washburn (1956, 1979) divided pat- ﬁ res and disturbance of tundra vegetation) terned ground based on (1) geometry to circles,

10 Periglacial phenomena

Figure 3. An example of thawing palsa in Biesjeaggi mire (69°35’39”N, 26°10’46”E / 355 m a.s.l. / 15th of August 2002). The marginal part of the permafrost mound has subsided and formed a small thermokarst pond.

Figure 4. Small ca. 10 m in diameter vegetated thermokarst pond in Biesjeaggi mire (69°27’20”N, 26°8’37”E / 345 m a.s.l. / 15th of July 2003). In the foreground a wind defl ated bare peat surface can be discerned. polygons, nets, steps and stripes and (2) sorting tion landforms (e.g. van Everdingen 2005: 53), to sorted and non-sorted features. Sorted pat- boulder depressions (Lundqvist 1962), taluses terned ground is defi ned by alternation of fi ne (Kejonen 1997) and even palsas (Meier 1987: and coarse soil material, whereas non-sorted 192). patterned ground is formed by microrelief Several hypotheses for patterned ground and/or vegetation differences (see Washburn formation have been proposed (e.g. Washburn 1956: 826–838). Some researchers have used 1970; Nicholson 1976; Mackay 1980; Ray et al. a broader defi nition that also includes solifl uc- 1983; Schunke & Zoltai 1988; van Vliet-Lanoë

11 Periglacial phenomena

Figure 5. Patterned ground, sorted polygon, in a temporary ponding depression close to lake Áhkojávri (69°34’57”N, 26°13’34”E / 375 m a.s.l. / 8th of July 2002). The diameter of the ﬁ ne soil centre in the middle is ca. one metre.

1991; Werner & Hallet 1993; Kessler & Wern- of the non-sorted circles tend to be dome- er 2003; Peterson & Krantz 2003; Matsuoka shaped (Fig. 6) and bare-soil circles may be et al. 2003; for reviews see Washburn 1956, cracked to non-sorted polygons. Circles are 1997) but it is still widely accepted that many predominant in different periglacial environ- of them are polygenetic in origin (e.g. French ments (Washburn 1979: 128). Several mor- 1996: 142; Mann 2003). For example at a local phological types of sorted and non-sorted scale, microclimate, topography, soil proper- circles exist. Earth hummocks (e.g. Nicholson ties, moisture availibility, vegetation cover and 1976; Mackay 1980; Schunke & Zoltai 1988), snow distribution are important factors (e.g. peat pounus (e.g. Salmi 1972; Seppälä 1998; Matthews et al. 1998). Furthermore, similar van Vliet-Lanoë & Seppälä 2002), stony earth features can be formed due to different proc- circles (e.g. Williams 1959) and mud or frost esses and the same process may produce dis- circles (e.g. Shilts 1978; Harris 1998; Peterson similar features (Washburn 1979: 157). et al. 2003; Walker et al. 2004) are different Circles, polygons and nets are features kinds of non-sorted circles (see van Everdin- whose mesh is dominantly circular, polygo- gen 2005: 53), although Washburn (1956: 830) nal or net-like. Sorted patterns have a border classifi ed earth hummocks as non-sorted nets. of stones surrounding fi ner soil material and Stone pits (e.g. Lundqvist 1962: 21–24; Seppälä non-sorted patterns are often margined by 1987: 48) and debris islands (Washburn 1956: vegetation and they lack a stony border. A 827) have a different appearance than normal furrow or a crack may delineate non-sorted stone-bordered sorted circles (Fig. 7). For ex- polygons and nets (Washburn 1956: 826–833). ample, Hallet (1998), Kessler et al. (2001), Hol- The main processes responsible for the for- ness (2003) and Haugland (2004) have studied mation of circles, polygons and nets are differ- sorted circles more recently. ential frost heave, mass displacement and/or The size of the non-sorted polygons varies frost sorting and, in addition, soil cracking for from fi ve centimetres to even 100 m and the polygons (e.g. Washburn 1970: Table 1, 1979: diameter of the sorted features can range from 160–170). ten centimetres to ca. ten metres (Fig. 5; Wash- The diameter of the circles ranges com- burn 1979: 133, 142). Non-sorted polygons monly from 0.5 to three metres. Central areas are predominant in the continuous permafrost

12 Periglacial phenomena

Figure 6. Convex non-sorted circles from the valley northeast of Guivi fell (69°37’35”N, 26°28’33”E / 500 m a.s.l. / 22nd of June 2002). The circles are 20–40 cm high and their diameter varies from two to three metres. The size of the backpack is 45 x 35 cm.

Figure 7. Sorted circle, debris island, in a blocky soil material in Čeavvresjohka valley (69°33’15”N, 26°25’14”E / 340 m a.s.l. / 6th of July 2002). The iron probe is ca. 80 cm long. regions but small features are also found in stances the initial process (e.g. Washburn 1956, areas of seasonal frost (e.g. Ballantyne & Mat- 1970; Goldthwait 1976; Nicholson 1976; van thews 1983). Sorted polygons occur in dif- Vliet-Lanoë 1991; Kessler & Werner 2003). ferent periglacial environments, respectively Nets are intermediate features whose (Washburn 1979: 133–145). In general, the mesh is neither dominantly circular nor po- processes forming the sorted polygons are of- lygonal because the formative mechanisms ten the same as in the other sorted patterned have been weak and/or mixed or the features ground but soil cracking is in several circum- have stretched on slopes. Otherwise, the sizes,

13 Periglacial phenomena

Figure 8. Boulder depression above the treeline in the valley southwest of Guivi fell (69°36’33”N, 26°20’8”E / 425 m a.s.l. / 18th of July 2004). Signs of activity such as recently heaved and frost shattered blocks can be seen in the area. processes and distributions resemble circles and the material fi ner downward. Features are and polygons (e.g. Washburn 1979: 146–147; formed by frost heave and sorting but boul- Ballantyne & Matthews 1983; Haugland 2004: ders can be split by frost wedging (e.g. Lun- 5). In general, nets have been described from dqvist 1962: 73–76). different periglacial environments (e.g. Dabski According to Washburn (1956: 833–834) & Gryglewicz 1998; Hodgson & Young 2001). steps are features with step-like form and However, confusion of the circles, polygons downslope border of vegetation or stones and nets is relatively common in the litera- embanking an area of relatively bare ground ture (Lundqvist 1962: 29; Ballantyne & Harris upslope (Fig. 9). Steps are usually derived from 1994: 195). circles, polygons, or nets rather than developed Sorted fi elds are areas with a pure boul- independently. Features are normally less than der material on the ground surface and they one metre high and locate on 5°–15° slopes form no characteristic patterns (Lundqvist (Washburn 1979: 149). The sorted steps have 1962: 61–63). However, they are separated stony riser and they are probably derived from from the frost weathered block fi elds (Chap- sorted circles or polygons. The non-sorted ter 2.2.5). Boulder depressions are the most steps have often vegetated riser and they are common type of the sorted fi elds and they are derived mainly from different hummocky non- mainly described from different parts of Fen- sorted circles. In general, studies on the step- noscandia (e.g. Lundqvist 1951; Ohlson 1964; like patterned ground are quite limited (Wash- Aartolahti 1969; Piirola 1969; Seppälä 1982b; burn 1969: 150–151; Walsh et al. 2003a). Söderman 1982; Hättestrand 1994). Boulder Stripes are features with a striped pattern depressions are fl at barren fi elds of pure boul- oriented down the steepest available slope. der material situated in shallow depressions in Sorted stripes are formed of parallel stony the landscape (Högbom 1905: 27). Features lines whereas non-sorted stripes are a set of are the most typical in forested areas but they vegetated and relatively bare ground lines (Fig. also occur above the treeline in the barren fell 10; Washburn 1979: 151–156). Stripes can areas (Fig. 8). Boulder depressions’ diameters be some hundred metres long and up to two vary from some metres to several hundred me- metres wide or even more. Features occur on tres and the small forms resemble stone pits. the slope gradients from two to ca. 30° and The largest boulders are typically at the surface rarely up to 40°. Hummocky stripes are one

14 Periglacial phenomena

Figure 9. Ca. 60 cm high non-sorted steps on a valley slope northwest of Suophášoaivi fell (69°30’6”N, 26°24’19”E / 430 m a.s.l. / 31st of July 2002).

Figure 10. Non-sorted stripes on a fell slope southeast of Uhc-Áhkováráš fell (69°32’41”N, 26°9’14”E / 370 m a.s.l. / 5th of July 2003). Width of the bare ground stripes are ca. 0.5 m. type of the non-sorted features where hum- (2003) have studied the formation of sorted mocks are lined up to form the stripes (e.g. stripes. Studies on the non-sorted stripes are Lundqvist 1962: 58–59). Stripes occur in many rather limited (for exceptions see Washburn environments and non-sorted have somewhat 1947: 94, 1969: 151–154; Lundqvist 1962: 58– the same distribution as the non-sorted circles 59; French 1974). and polygons although stripes appear to be less common (Washburn 1979: 151–156). For example, Ray et al. (1983), Muir (1983), Werner and Hallet (1993), Hall (1994), Francou et al. (2001), Holness (2001) and Matsuoka et al.

15 Periglacial phenomena

2.2.4 Solifl uction and other slope phenomena tant environmental determinants for the processes are soil moisture, slope gradient, grain size distribution and vegetation cover (Fig. 1; Periglacial slope processes can be subdivided Washburn 1979: 198–204; Pissart 1993; Mat- to two main types (1) solifl uction and (2) more suoka 2001a). Solifl uction produces diverse set localised rapid slope failures that occur spo- of landforms such as sorted and non-sorted radically often during thaw season. Solifl uc- sheets, lobes and terraces (e.g. Rapp 1960; tion is a slow downslope movement of soil Washburn 1967; Benedict 1970; Harris 1987, mass usually associated with freeze-thaw cy- 1996; Lewkowicz 1988; Matthews et al. 1993; cles and frost heave (Andersson 1906: 95–96; Matsuoka 2004). Matsuoka 2004: 984). The second main sub- Solifl uction sheets are relatively smooth type of the mass wasting phenomena (rapid and continuous debris mantles that cover large mass movements) contains a variety of falls, areas up to several square kilometres (Wash- avalanches, fl ows and slides (e.g. McRoberts & burn 1979: 214; Ballantyne & Harris 1994: Morgenstern 1974; Washburn 1979: 192–197; 205–212). Non-sorted solifl uction sheets are Innes 1983; Nyberg 1985; Ballantyne & Harris partly or totally vegetation covered and sorted 1994: 118). sheets have stony riser or they may be block- Solifl uction can be considered as a collec- covered (Fig. 11; Ballantyne & Harris 1994: tive term, which includes frost creep result- 209). Washburn’s (1979: 219) block slopes and ing from nearly vertical settlement of soils partly block fi elds are similar to the boulder heaved normal to the slope and gelifl uction sheets and can be included into this class (cf. representing downslope displacement of ice- Dahl 1966). In solifl uction sheets, the whole rich soil during thawing (e.g. Washburn 1979: slope is in a rather uniform movement and 201). Both frost creep and gelifl uction operate they may terminate dowslope in regular risers together and the term solifl uction is typically i.e. solifl uction terraces (Matsuoka 2004: 986). used to describe their combined effects (Mat- Solifl uction sheets are generally more com- suoka 2004: 984). Solifl uction may operate even mon in gently sloping polar regions but they on one degree slopes and the annual rates of can also be found in mountain areas (e.g. Fran- downslope movements varies between ca. 0.5 cou & Bertran 1997). Topographically, sheets and 10 cm a-1 (Matsuoka 2004: 984). Impor- locate typically at upper slope areas (Ballan-

Figure 11. Sorted soliﬂ uction sheets on the eastern and southeastern slopes of Suobbatoaivi fell (69°38’5”N, 26°12’10”E / 430–500 m a.s.l. / 14th of August 2002).

16 Periglacial phenomena tyne & Harris 1994: 205). According to French been described from different periglacial envi- (1996: 156) they are the least studied but the ronments (French 1996: 156). After Matsuoka most widespread solifl uction landforms. (2004: 986) the lobes are the most widespread Solifl uction terraces, also termed benches solifl uction features and non-sorted are more by Washburn (1979: 214–215), are up to three common than sorted ones (see Benedict 1976: metres high terrace-like landforms often situ- 60). The solifl uction lobes have been studied, ated in valley-fl oor or in other similar topo- for example, by Nesje et al. (1989), Yamada et graphical locations where the slope gradient al. (2000), Hugenholtz and Lewkowicz (2002), decreases (Ballantyne & Harris 1994: 115, Matsumoto and Ishikawa (2002) as well as Jae- 205–206). The longest dimension of the ter- sche et al. (2003). races tends to be parallel to the slope and they Solifl uction streams are narrow linear de- develop on the slopes or foot of the slopes posits of debris or coarse block material where the movement rates are relatively uni- (Washburn 1979: 216). Most of the described form along the contours. Vegetation, which solifl uction streams have been sorted block has a restraining effect upon movement, can streams (e.g. Andersson 1906: 97–104; Wash- assist the soil thickening and terrace growth burn 1979: 219; Åkerman 1980: 105; Tyurin (Benedict 1976: 60; Washburn 1979: 214–215). 1983; Harris et al. 1998; Boelhouwers et al. Features have been under investigation, for 2002; Boelhouwers & Sumner 2003). Instead, example, by Dongxing et al. (1993) in Qinhai- the descriptions of non-sorted streams are Xizang Plateau, China. In general, non-sorted rather limited (e.g. Washburn 1947: 88–96). solifl uction terraces are more widely distrib- Ploughing blocks, also termed ploughing uted and more common than sorted features boulders, are a kind of solifl uction feature (Benedict 1976: 60). consisting of isolated, commonly boulder- Solifl uction lobes have a tongue-like ap- sized stones that leave a linear trough upslope pearance with relatively steep frontal riser up and form a low ridge downslope (Fig. 12). The to 1.5 m high (Washburn 1979: 216; Matsuoka size of the depression and ridge are directly 2004: 986). Lobes occur often on relatively connected to the size of the block (Washburn steep mountainous regions where the fl ow is 1979: 223; Hall et al. 2001: 223). Ploughing channelized (Benedict 1976: 60) but they have blocks have been described mainly from the

Figure 12. Ploughing block with downslope ridge west of Guivi fell (69°37’14”N, 26°22’25”E / 480 m a.s.l. / 19th of June 2002). The backpack is 45 cm high.

17 Periglacial phenomena alpine region (e.g. Tufnell 1972; Reid & Nesje discontinuous and relatively slow but more of- 1988; Ballantyne & Harris 1994: 216–218; Bal- ten it is rapid slope process. Resulting deposits lantyne 2001; Berthling et al. 2001). By con- are fan-like and the material is chaotic ranging trast, with ploughing blocks, which move fast- from fi ne sand to boulders (Washburn 1979: er than the surrounding material, blocks which 195). Heavy rainfalls and quick temperature impede solifl uction are termed braking blocks rise in the late winter usually trigger slushfl ows (Fig. 13). Soil, stones and vegetation can be (e.g. Nyberg 1985, 1989; Onesti 1985; Clark & piled up on their upslope side and a depres- Seppälä 1988; Elder & Kattelmann 1993). sion can be created on the downslope side of Talus is an apronlike accumulation of the block (Åkerman 1980: 105). rockfall debris that can be many metres thick Debris fl ows are rapid mass movements of (Washburn 1979: 231). The term talus is used water saturated (10–50%) soil material (Innes to describe both the slope form and its mate- 1983: 469–470, van Everdingen 2005: 16). For rial (Ballantyne & Harris 1994: 219). Taluses example, Brunsden (1979) has classifi ed them occur in different mountainous environments into catastrophic, hillslope and valley-confi ned but they are most common in periglacial refl ows. Extreme rainfalls of high intensity trig- gions (e.g. Rapp 1986; Albjär et al. 1979). They ger debris fl ows during the summer or autumn are attributed to weathering and falling of (Ballantyne & Harris 1994: 118). They often rock material from the rock faces (Fig. 14). start as a debris slide but quickly the move- The rock material in the talus has fall sorting, ment changes from sliding to viscous slurry that is, the largest blocks rest at the base of the fl ow. The resulting debris fl ow track can be talus and the smaller blocks and stones cover ribbon-like, fl anked by bouldery levées and the middle and top parts of the formation. If terminate in a few metres to 20 m wide lobate the cliff face is relatively low there can also front (Rapp 1986: 59–60). Debris fl ows occur exist reverse sorting. Stratifi cation of the talus in mountainous environments in both perma- tends to be poor to absent (Washburn 1979: frost and nonpermafrost areas (e.g. Larsson 231–234). Talus sheets, talus cones and coa- 1982; Nyberg 1985). lescing talus cones are different type of talus Slushfl ows are avalanches of water-satu- landforms. Secondary processes such as snow rated snow combined with debris (Washburn avalanches, slushfl ows, debris fl ows and pro- 1979: 193; Rapp 1986: 62). Slushfl ow can be talus rock glacier formation can modify talus

Figure 13. Large (one metre high and over two metres wide) block preventing the soil creep on the slope of Guivi fell (69°37’18”N, 26°24’33”E / 515 m a.s.l. / 18th of July 2004).

18 Periglacial phenomena

Figure 14. Talus features under rock faces in the Geavvu canyon northeast of Ruohtir fell (69°29’20”N, 26°29’7”E / 335–375 m a.s.l. / 31st of July 2002). slopes. Ballantyne and Harris (1994: 219–244) In general, block fi elds can be divided into have made a comprehensive review of talus autochthonous that consists mainly of the phenomena. Among others Hinchliffe et al. products of in situ weathered bedrock, and (1998), Hétu and Gray (2000), Sass and Wollny allochthonous, in which boulders have some (2001) as well as Curry and Morris (2004) have other origin, for example glacial (Ballantyne studied taluses more recently. & Harris 1994: 173). The periglacial block fi elds are surfi cial layers of moderate-sized or large angular shattered rocks formed in 2.2.5 Periglacial weathering cold environments (Washburn 1979: 219; van features Everdingen 2005: 6–7) and they are mainly autochthonous. According to Rudberg (1977: Weathering processes are probably more com- 94) blocks should cover over 50% of the land plicated than previously thought in periglacial surface in the block fi elds. Block slopes are not environments (e.g. Ballantyne & Harris 1994: treated here because they have distinct con- 163; French 1996: 7, 40–45). Furthermore, the nection to frost creep and gelifl uction process- role of chemical denudation may be under- es and were described among the solifl uction estimated (e.g. Rapp 1960; Thorn et al. 2001). landforms (cf. Washburn 1979: 219). Impor- Traditionally researchers have concentrated tant processes in the formation of the block on frost-weathering (McGreevy 1981; Lau- fi elds are mechanical frost-weathering, elu- tridou & Ozouf 1982; Matsuoka 2001b; Hall viation of fi nes, frost sorting and solifl uction et al. 2002), that is disintegration and break- (Fig. 1; Washburn 1979: 219–221). However, up of soil or rock material by the combined the age and genesis of many extensive fi elds action of frost shattering, frost wedging and around the world is problematic (for summary hydration shattering (van Everdingen 2005: see French 1996: 236). Dahl (1966), Kleman 30). However, the exact mechanics of frost and Borgström (1990), Nesje and Dahl (1990) weathering are still incompletely understood and Dredge (2000) among others, have studied despite several hypotheses have been pre- block fi elds. sented (for summary see Ballantyne & Harris A tor is a residual mass of bare bedrock that 1994: 163–165). Herein the focus will be on rises above its surroundings, is isolated by free- the landforms produced by different weather- faces on all sides, and owns its formation to ing processes. differential weathering and mass wasting (e.g.

19 Periglacial phenomena

Figure 15. Ca. two metres high tor on summit close to Suobbatoaivi fell (69°38’29”N, 26°10’27”E / 475 m a.s.l. / 25th of July 2003).

Linton 1955: 476). Sizes of tors range from a Nivation is not a single process but rather a few metres to tens of metres and they occur collective term describing the erosional proc- often in upland areas on summits and slopes esses related to the snow patches or snow- (Fig. 15; Washburn 1979: 78). Different gen- banks (Thorn 1988: 10–11). It is the com- esis for the formation has been suggested (e.g. bined action of intensive freeze-thaw activity, Linton 1955; Pullan 1959; Demek 1964; Dahl enhanced chemical weathering, slopewash and 1966; Selby 1972). Proposed processes include transport of debris by solifl uction (Thorn frost action, mass wasting, wind action, and 1988: 11–17; cf. Fig. 1). In general, the zone sub-surface weathering and later exhumation of effective erosion is around the edges of a (Fig. 1; Washburn 1979: 78). Tors may have snow patch where meltwater is available but various origins and may evolve under different the ground is not protected with insulating climate conditions (French 1996: 236–238). In snowcover against atmospheric freeze-thaw general, tors are common in periglacial areas cycles (Embleton & King 1975: 130). Further but tors cannot be kept as clear indicators of information on the problematic nivation phe- periglacial environment (e.g. see Embleton & nomena can be obtained from Lewis (1939), King 1975: 174–175; French 1996: 238). Thorn (1979, 1988), Thorn and Hall (1980, 2002) and Christiansen (1998). Thorn (1988: 26) recommended not to use the term nivation 2.2.6 Nival phenomena because it is rarely possible to assess the degree to which nivation has modifi ed the location of Nival phenomena are often associated only a contemporary snow patch, however, it has with the erosive nivation process despite the still been relatively widely used in the scientifi c fact that depositional processes and landforms literature (e.g. Ballantyne & Harris 1994: 245). such as niveo-eolian phenomena (e.g. Wash- Nivation [or snow accumulation (Thorn burn 1979: 266–267) and protalus ramparts 1988: 27)] hollows are shallow depression- or (e.g. Washburn 1979: 234–235; Ballantyne & step-like features often ranging from a few Harris 1994: 236–240) also exist. However, metres to several tens or even hundreds of depositional phenomena are not treated here metres length and width (e.g. Cook & Raiche because of their infrequency in Fennoscandia 1962; Embleton & King 1975: 135; Ballantyne (e.g. Seppälä 1987). 1978). Nivation hollows develop often along

20 Periglacial phenomena

Figure 16. Large over 600 m long longitudinal nivation hollow-like snow accumulation site north of Čeavrresgielas fell (69°35’18”N, 26°22’29”E / 420–460 m a.s.l. / 5th of June 2002). lee slopes where the snow is forming large sizes of the features range from a few metres snowbanks (Embleton & King 1975: 135). to some tens of metres. Common occurrence Lewis (1939: 153) recognised three different areas are on valley fl oors, near lakes or streams, nivation hollow types: transverse, longitudi- and below breaks of slopes in different per- nal, and circular. Transverse hollows (nivation iglacial environments (Embleton & King benches) are step- or terrace-like features cut 1975: 143; Washburn 1979: 173). The origin is into the hillside. They are the most common incompletely understood but supposed proc- nivation features (Åkerman 1980: 106). Ter- esses responsible for the formation are up- race-like nivation hollows can be initial stages freezing of stones, ground saturation and re- of cryoplanation terraces (e.g. Priesnitz 1988). moval of fi nes by meltwater, the rotation and Longitudinal nivation hollows follow the di- shifting of the stones in the saturated ground rection of maximum ground slope and they under their own and overlying snow and icing are generally associated with small stream weight (Washburn 1979: 173). For example, courses (Fig. 16; Embleton & King 1975: 137; Mackay and Mackay (1976) and Christiansen Åkerman 1980: 108). Circular nivation hol- (1998) have studied stone pavements. lows are largely independent of structural or water-eroded features and they occur more often on gently sloping ground. The diameter 2.2.7 Aeolian processes and of the circular features can range from a few landforms tens of metres up to one kilometre. The largest circular hollows are called nivation cirques Aeolian activity produces both erosion and (e.g. Watson 1966) and resemble real cirques depositional features in cold climates (e.g. formed by glaciers. A continuum between Niessen et al. 1984; Seppälä 2004b). Defl ation large snow patch hollows and glacial cirques depressions (Fig. 17) and surfaces, ventifacts probably exists (Embleton & King 1975: 138, (e.g. Schlyter 1995), pavements (e.g. Schönhage 142–143; Thorn 1988: 22–23). 1969), wind patterned ground and turf-banked Stone pavements consist of closely packed terraces (for review see Ballantyne & Harris stones or boulders whose fl at surfaces are up- 1994: 261–267) are different types of erosion permost and at a level with each other. The features. Deposition processes can form fea-

21 Periglacial phenomena

Figure 17. Defl ation depression formed in a sand dune on the eastern slope of Geatgielas fell (69°27’2”N, 26°22’22”E / 410 m a.s.l. / 24th of June 2003). tures such as sand dunes (e.g. Carter 1981), the defl ation basins can be stone pavements cover sands (e.g. Maarleveld 1960) and loess- (cf. Chapter 2.2.6; Seppälä 2004b: 174). Wind like silts (e.g. Péwé 1955; Pye 1984). Herein can also erode peat deposits and features such the focus will be on the defl ations and dune as palsas (e.g. Fig. 4; Seppälä 1972b; Luoto & landforms and more about aeolian phenom- Seppälä 2000; Seppälä 2003a). ena in periglacial regions can be obtained from In periglacial regions, sand dunes have been Åkerman (1980: 228–257), Koster (1988) and mainly formed during Pleistocene deglacia- Seppälä (2004b). tion and are inactive and stabilized nowadays Wind defl ation is the blowing out of sand (Embleton & King 1975: 191). However, in and fi ner particles from the ground cover and some areas, dunes are still developing (e.g. their transportation by the wind (French 1996: Seppälä 2004b: 197–206). The sizes of the 206). The resulting features are different sized dune ridges range from less than one metre to defl ation surfaces and depressions (Fig. 17). 30 m high and up to several hundreds of me- The defl ation surfaces are from less then one tres or even kilometres long. Furthermore, the metre to tens of metres wide and only a few dune areas range from less than one hectare to centimetres deep bare ground areas, whereas several thousands of square kilometres (Wash- the defl ation depressions are from less than burn 1979: 264–265). In general, the most one metres to even over ten metres deep and common dune type is parabolic where arms up to hundreds of metres in diameter (e.g. are pointing away from the direction of move- Seppälä 1995b: 799; French 1996: 206). Defl a- ment (e.g. Seppälä 1971, 1972c; Carter 1981; tion of the fi ne sand deposits usually ceases Koster 1988). Linear features (transverse and when the level of ground water or underlying longitudinal) are rather seldom. ground moraine is reached. At the bottoms of

3 STUDY AREA

3.1 Location and topography ites and gabros (Meriläinen 1965). Present re- The study area is located in the northernmost lief of the study area has formed by tectonic Finnish Lapland over 300 km north of the block movements and denudation phases dur- Arctic Circle close to the Norwegian border ing the last 25 million years. The fell groups (SW corner 69°22’39”N, 25°58’46”E and NE of the Báišduattar and Áilegas are two of the corner 69°38’55”N, 26°28’54”E) (Figs. 18 & seven horsts that have been lifted up by block 19). The cover of the study area is 600 km2 (20 movements in the Tertiary 25–10 million years x 30 km). The topography of the Báišduattar ago (Mikkola 1932: 31–34; Tanner 1938: 218). – Áilegas region is characterized by gently Afterwards, the tectonic blocks have split and sloping fells (Figs. 19 & 20). The fell sum- erosion has rounded them into separate fells mits are relatively fl at and rounded although and groups of fells. Some of the split lines can tors are found in many places (Kaitanen 1969, be seen in the present landscape as high angle 1989; Fig. 15). In general, valleys between the shear zones and fracture lines (Fig. 20; Geo- fells and groups of fells are wide and fl at-bot- logical map 1:1 000 000 1987). tomed, but there are also some relatively steep- sided valleys such as the polygenetic Geavvu canyon in the east (see Figs. 14 & 19). Small- 3.3 Weichselian glaciation, scale topographical variation prevails in the deglaciation and soil types valleys where glaciofl uvial accumulations are abundant and on the slopes where the glacial Glaciers in the Weichselian have shaped the melt water channels dominate (Figs. 20 & 21). general topography of the area and formed Elevations in the area range from 110 to surfi cial deposits during the last twenty thou- 641 m a.s.l. with a mean of ca. 370 m calculated sand years. However, the glaciers eroded the from the digital elevation model (see Chapter bedrock only slightly because of three main 4.1.3). The highest fells are Guivi (640.5 m reasons (Kaitanen 1989: 7–10). Firstly, the a.s.l.), Guovdoaivi (ca. 620 m a.s.l.) and Lánká study area is located in an interlobate zone be- (619.8 m a.s.l.) (Fig. 19). Furthermore, there tween the Finnmark and Tuloma ice streams are about 15 other over 550 m high fells mainly where the ice fl ow was very weak (Punkari located in the northeastern part of the region. 1996: 16). Secondly, Kaitanen (1989: 7) esti- Most of the study area is in a relative narrow mated that the ice thickness could not exceed altitude zone, because over 60% of the area is 1200 m even during the glacial maximum located between 300 and 400 m a.s.l. Relative and the ice cover was mostly less than 600 m relief varies locally from less than ten to more thick during the other glacial phases. Thus, the than 200 m. thickness of the ice on the fells has generally been less than 400 m. Thirdly, Kaitanen (1989: 7) also suggested that the ice could have been 3.2 Bedrock and general cold based, which would prevent basal erosion geology (cf. Hättestrand & Stroeven 2002). Altogether, the weak erosion of the continental ice sheet Geologically, the study area belongs to the is clearly demonstrated by the presence of nu- Pre-Cambrian ca. 1.9 billion years old granulite merous preglacial weathering remnants (e.g. complex (Mikkola 1941; Meriläinen 1976). The Kaitanen 1969; Fig. 15). main rock types are fi ne- and coarse-grained The deglaciation of the Báišduattar – Áile- garnet-quartz-feldspar gneisses, but there ex- gas area started ca. 10 000 years BP (uncalibrat- ist small areas of quartz, diorites, granodior- ed). The ice margin retreated approximately

23 Study area

Figure 18. Location of the study area in northern Fennoscandia. In the index map the dark grey represents permafrost areas (continuous, discontinuous, sporadic and alpine; after Harris 1986: 2) and lineation indicates the ice sheet of Green- land.

60–250 m a-1 toward the southwest and the fi nally released from under the ice ca. 9850– ice thinning ranged from one to fi ve metres 9800 years BP (Seppälä 1971: 61, 1980: 318). per year depending on the deglaciation stage Glacigenic till is the predominant soil type (Seppälä 1980: 319; Koskinen 2005). The local covering roughly 80% of the study area but topography had a signifi cant effect on the de- this estimate includes blocky surfi cial material glaciation pattern. Therefore, the retreat rate as well. On the fells and slopes the moraine of the ice could have been less than generally layer is thin (0–3 m) but at the bottoms of the estimated in the fell areas. Towards the end of valleys its depth may increase up to ten metres the deglaciation the ice mass became more or (Syrilä 1964). Block fi elds are more frequent less stagnant due to the thin ice cover and rug- on the fell slopes and above the treeline but ged topography. At fi rst, the summits of the they are not, like in many parts of Northern fells were released under the continental ice Finland, abundant on the fell summits (e.g. Fo- and were subject to frost processes in severe gelberg & Seppälä 1986). The blocks are gen- periglacial conditions. When about half of the erally relatively rounded and have commonly study area was deglaciated ice-dammed lakes a glacigenic origin although also small frost formed in the central parts of the area (Kos- shattered block fi elds exist in the area. Most kinen 2005: 48). They existed probably only of the glacigenic deposit is basal till but there from some years to at the most a few decades are some areas in the northeast and south were (e.g. Seppälä 1980). The whole study area was the hummocky ablation moraine dominates.

24 Study area

Sand and gravel deposits, which cover ca. cial and glaciofl uvial processes. 7% of the area, are mainly located in the valleys In general, glacial landforms are rare due to where the glaciofl uvial action was predominant weak glacial erosion and unfavourable depo- during the deglaciation. At some locations, sitional conditions. However, pre-Weichse- fl uvial and aeolian processes have affected lian cirque-like bedrock landforms (Kaitanen the glaciofl uvial deposits and formed second- 1969: 46) and a few large drumlinoids occur ary sand and gravel deposits. Sand dunes can in the study area (Map of Quaternary geol- be found in the southeast and fl uvial deposits ogy 1:1 000 000 1988). Kettle holes that can in the major river valleys in the northern and be over 20 m deep are rather common glacial northeastern part of the study area (Fig. 19). landforms. Furthermore, hummocky moraines Three considerable esker systems (see next dominate relief variability in some areas, for chapter 3.4) and some smaller glaciofl uvial example near the Sieiddejávrrit and Basijávri landforms such as kamehummocks occur in lakes. More detailed description of the textural the region. In general, the material of the gla- and structural characteristics of the glacial ciofl uvial deposits is sorted relatively poorly. depositional landforms is given by Karzewski The organic deposits are prevalent in the (1975). wide fl at-bottomed valleys in the central parts Glaciofl uvial action was very intensive dur- of the study area. The total cover of peat is ing the deglaciation (e.g. Seppälä 1980). Stag- about 11%. Peat layers are generally thin (less nant ice produced a huge amount of melt than one metre) but in the extensive palsa mires, water that grooved a number of subglacial, for example in Biesjeaggi and Čulloveijeaggi marginal and extramarginal channels into the mires, the deposits can be up three metres moraine cover and even into the surface of thick (e.g. Lappainen & Hänninen 1993). Silty bedrock (Koskinen 2005). The most abun- soils are rather common in the valleys between dant glaciofl uvial channel types are marginal Áitečohkka and Stuorra Biesvárri fells. The and submarginal, which predominant on the silty sediments deposited at the bottoms of fell slopes (Fig. 21). At the higher altitudes, be- the temporary ice-dammed lakes during the tween the fell summits, occur many overfl ow deglaciation (see above) and are often covered channels that may continue as subglacial chan- by organic material at present. nel into the valleys. Extramarginal channels are in the minority and they usually locate on outwash plains. 3.4 Geomorphology of the Glaciofl uvial deposition landforms are Báišduattar – Áilegas relatively abundant but their distribution is limited to the valleys (e.g. Seppälä 1993a: 270). General relief of the study area is governed Luopmošjáguotkku esker in the southern part by bedrock topography, as usual in Northern is a typical 10–30 m high steep-sided mid-val- Finland (Fogelberg & Seppälä 1986), although ley esker (Fig. 22). It is oriented from southwest long term fl uvial erosion has modifi ed the to northeast representing the general retreat landscape signifi cantly in the northwest (Figs. direction of the ice edge during the deglacia- 19 & 20). The medium-scale topographical tion. Kettle holes and kamehummocks (5–25 variation is mainly caused by glaciofl uvial ac- m high) dominate most of the marginal areas, tion and most of the small-scaled landforms which are often better sorted than the esker are frost formed (cf. Seppälä 2005b). Bedrock ridges. In the middle part of the study area landforms are not very abundant but tors are there is situated a discontinuous and ca. 15–35 found on many fell summits (e.g. Fig. 15; Kai- m high esker system that is mainly southwest tanen 1989: 9). In addition, a few variously northeast -oriented. The northernmost west oriented fracture lines and valleys occur in east -directed Goike-Sitnogohpi esker has the area (Fig. 20). The most signifi cant is the more undulating topography than the previ- Geavvu canyon, which is also modifi ed by gla- ous resembling kame-esker. The hummocks

25 Study area

Figure 19. Generalised map of the study area. The forest and mire data are obtained from biotope database (© Metsähallitus 2006). The vertical interval of contours is 20 m (© National Land Survey of Finland, Licence number 49/MYY/06).

26 Study area

Figure 20. Three-dimensional terrain view of the study area. The terrain view was produced using the shaded relief surface model derived from the created digital elevation model (sun angle = 45°, azimuth = 315°) (Chapter 4.1.3).

27 Study area

Figure 21. Lateral glacial melt-water channels on the northwestern slope of Gaskkamušaláš fell (69°26’35”N, 26°2’31”E / 530–560 m a.s.l. / 31st of July 2003).

Figure 22. Luopmošjáguotkku esker between the Stuorrajávri (323 m a.s.l., on background) and Nuorttatjávri (324 m a.s.l.) lakes in the southern part of the study area (3rd of August 2003). The photograph is taken from the northeast looking southwest. are from fi ve to 25 m high and the tops are has grooved over 100 m deep river valleys to usually defl ated. In addition, kame-landforms, the northwest (Figs. 19 & 20). At present, the glaciofl uvial deltas, outwash plains and small Áhkojohka and Čulloveijohka rivers have a valley trains exist in the study area. minor role in the relief alteration despite the The ice-dammed lakes in the central and fact that spring fl oods can be very drastic (see southern parts of the area have left several Chapter 3.7). shore marks commonly between the altitudes Sand dunes and defl ations represent aeolian of 330 and 360 m a.s.l. (Seppälä 1993a; Kos- landforms of the study area. The dunes were kinen 2005: 48). The long-term fl uvial action formed relatively soon after the deglaciation, has shaped the topography clearly more than whereas defl ation is the most signifi cant aeo- the shore processes. Preglacial fl uvial action lian process nowadays (e.g. Seppälä 1971). The

28 Study area sand dunes are located in the southeastern (1994) studied the present and ancient solifl uc- corner east of the Luopmošjáguotkku esker tion in the Áilegas area using different types that offered material for wind transportation. of measuring lines, a laseroptical geodimeter, The dunes are mostly parabolic and longitu- block towers dug in the ground, pollen analy- dinal indicating the direction of the affective sis and radiocarbon dating. Luoto and Sep- winds from the northwest (Seppälä 1971: 15, pälä (2002a, 2003) have modelled palsas and 1993a). Wind has not only eroded the dunes thermokarst ponds and their study area in- and glaciofl uvial landforms but also morainic cluded the Báišduattar – Áilegas area. Seppälä hummocks. Seppälä (1971) gives a detailed (2003a) studied surface abrasion of palsas near description of the aeolian processes from an Lake Áhkojávri. In the same area Rönkkö and area about 20 km southeast of the Báišduattar Seppälä (2003) investigated the active layer of – Áilegas area. palsas using fi eld measurements, surface char- Biogenic processes have mainly levelled acteristics and statistical analyses. Hjort and the general topography but together with the Seppälä (2003) described briefl y topographical frost action they have generated small-sized chatacteristics of active and inactive patterned cryogenic hummocks (van Vliet-Lanoë and ground in the northeastern part of the present Seppälä 2002) and permafrost cored palsas. study area. Moreover, Luoto and Hjort (2004, At present, the frost phenomena are one of 2005, in press) as well as Hjort and Luoto (2005, the most effi cient processes that modify the 2006) modelled with different statistical tech- land surface in the study area. The main proc- niques the distribution of patterned ground in esses on the level ground are cryoturbation the same region. and frost sorting and on the slopes solifl uc- Most of the periglacial studies performed tion. The human infl uence on the topography near the Báišduattar – Áilegas have been car- has been very limited throughout history. The ried out in the Skallovarri – Vaisjeaggi area most widespread but indirect modifi cation some 45 km to northeast from the present has been reindeer husbandry. The area is lo- study region. For example, Seppälä (1982a, cally overgrazed (Heikkinen & Kalliola 1989), 1990, 1994b, 1995a, 1998, 2003b, 2005c) has which has enabled small-scale water and wind made numerous fi eld experiments with pounus erosion. Reindeer may also affect the forma- and palsas in the area. In addition, the stud- tion of non-sorted patterned ground (Chapter ies of van Vliet-Lanoë and Seppälä (2002) as 6.2.3). well as Luoto and Seppälä (2002b) focused on pounus. Some investigations have been carried out in the Muotkatunturit (Muotkeduoddara) 3.5 Previous periglacial and Kaamasjoki (Gámasjohka) – Kiellajoki research in the study region (Giellasjohka) regions, which are located south and southeast of the study area. For example, Previous periglacial studies in the Báišduattar Piirola (1969, 1972) studied patterned ground – Áilegas area have concentrated on the spe- and Kejonen (1979) solifl uction phenomena cifi c processes or landforms and no one has in the Muotkatunturit fell area, and Seppälä performed an extensive regional geomorpho- (1971) aeolian processes in the Kaamasjoki- logical mapping. Söderman (1980) studied Kiellajoki river basin. Furthermore, Ruuhijärvi mainly slope processes in North Finland and (1962) has investigated palsas in Petsikko ca. 25 he had one study area located in the fell group km to the east and Hietaranta and Liira (1995) of Áilegas. King and Seppälä (1987) investi- as well as Liira and Hietaranta (1998) who have gated permafrost distribution in Northern studied talus slopes in the Geavvu river valley Finland with geoelectrical soundings and they close to Kevo research station. Vorren (1967), had three experiment sites in Áilegas. Seppälä Svensson (1971), Åhman (1977), Jahn and Sie- (1993a) has described the properties of sand dlecki (1982), Malmström and Palmer (1984) dunes of the southwestern parts. Kejonen as well as Meier (1987, 1991) are examples of

29 Study area preglacial studies conducted in northern Nor- have been 1.5–3ºC and treeline 200 m higher way fairly close to the present study area. than at present (Eronen 1979: 108–110; Seppä & Birks 2002: 195–197; Kultti 2004: 24). Be- tween 5000 and 4500 BP the temperature and 3.6 Past and present climate treeline started to degrade (e.g. Kultti 2004: 24–25) and the climatic conditions were again Northern Fennoscandia is climatically a unique more favourable for the periglacial processes area when compared to the other areas at cor- (Kejonen 1997: 103). Besides the climatic cool- responding latitudes. The climate is relatively ing, the lowering of the treeline was also attrib- mild due to the Gulf Stream that brings heat uted to the glacio-isostatic land uplift that has energy from the Gulf of Mexico to the North been about 60 m after the climatic optimum in Atlantic. Part of this energy is transported to the study area (Kejonen 1994: 53). During the Northern Fennoscandia by cyclones and west- last 4000 years climate has been relatively cool erly winds (Arvola 1987). Light and radiation with some colder stages, for example, between conditions at the study region are typical of 2900 and 2100 BP (Kultti 2004: 26) and during high-latitude areas, respectively. For example, the Little Ice Age ca. 150–400 years ago (e.g. the sun does not rise above the horizon be- Karlén 1976; Matthews & Briffa 2005). tween 26th of November and 15th of January The present climate of the study region is and the sun does not set between 17th of May subarctic (Seppälä 1976a) and relatively conti- and 26th of July. nental despite the distance to the Arctic Ocean Climamorphologically the region belongs is only about 100 km (Tuhkanen 1980: 78). to the zone of discontinuous permafrost (Sep- Climatological characteristics of the study re- pälä 1997b). Based on the geoelectrical sound- gion are presented based on the measurements ings, permafrost is probably widespread at the conducted at the Kevo Meteorological Station altitudes above 500 m and it can be up to 50 (69°45’N, 27°01’E), which is located about 35 m thick (King & Seppälä 1987, 1988). In the km apart from the study area to the north- valleys, only sporadic permafrost is present. east (Fig. 18; Table 6). It is notable that the A short review of the climate history of the climate parameters are measured in a forested Holocene is made before a more detailed de- valley at 107 m a.s.l. and may give too gener- scription of the present climate of the study alised a view of the climate of the Báišduattar area is given, because periglacial processes act – Áilegas area (e.g. Seppälä & Hassinen 1997). over a wide time span (Washburn 1979: 15) Therefore, climatological features of fell areas and the landforms can be stabilized and again are also discussed brieﬂ y. reactivated because of regional or global cli- The following meteorological data are mate change. taken from Climatological Statistics in Fin- The most favourable condition for the per- land 1961–1990 (1991) and Atlas of Finland iglacial processes was after the deglaciation (1987), if not otherwise mentioned. The mean ca. 9800 years BP before the area was cov- annual air temperature (MAAT) was –2.0°C ered by vegetation. However, this period was during the period 1962–1990 and it has var- quite short but still long enough to produce ied between +2.3°C and –6.7°C. The warmest distinct marks of frost action (e.g. Kejonen month, on average, was July (12.7°C) and cold- 1997; cf. Cook-Talbot 1991). The relatively est January (–15.7°C). The lowest measured cold climate phase lasted until 8700–8600 BP temperature was –47.9°C (1966) and the high- and it is estimated that the mean air temperate est was +32.9°C (1988). Mean of the freeze could have been ca. 2ºC lower than at present index was –2052.7 (std = 393.4), mean of the (Seppälä 1971: 73–79). The colder period was thaw index was 1372.7 (std = 108.2) and mean followed by a long (8300–4000 BP) warmer of the total temperature sum was –681.6 over phase with the Holocene thermal maximum at the period 1980–1991 (for more details see ca. 8000–6000 BP when the temperatures may Seppälä & Hassinen 1997: 154–155). In gener-

30 Study area al, these values indicate the presence of rather highest in July-August when about 31% (122.5 diverse set of periglacial phenomena (Harris mm) of the total rain amount of the year fell. S.A. 1982a). The driest season was February-May when the On average, there are 230 frost days (maxi- mean precipitation varied between 17.4 mm mum temperature of day < 0°C) in a year. The and 23.2 mm per month. On average, there mean number of the frost days at ground level was 173 days when snowfalls occurred and (minimum temperature at ground level < 0°C) only July and August were without any snow- is 258 and the ground level frost can occur in fall. Permanent snowcover develops after the summer months too. October and May are the end of October and it will last on open areas most important months when the freeze-thaw generally until 20th of May and in forests until –cycles are considered since at these times 25th of May (on average 210 snowcover days). temperatures ﬂ uctuate on either side of 0°C The mean maximum thickness of snow was all the time (Seppälä 1976a: 3). According to 69 cm. Liira and Hietaranta (1998: 4), there has been The mean annual wind speed was 2.9 m/ on the average ca. 80 freeze-thaw –cycles per s and the percentage of calms was 12. The year. Frost begins to form between the end of predominant wind direction was southeastly, September and beginning of October. Maxi- 29% of the winds with a mean of 2.7 m/s, but mum soil frost depth in till ground is about the strongest winds blew from the northwest 210–220 cm at sites cleaned of snow and in (mean = 4.6 m/s). The winds are, in general, open snow covered areas ca. 130 cm, but frost stronger in winter and the highest monthly can occur considerably deeper when snow mean is obtained in February (5.3 m/s). How- cover is thin (cf. Seppälä 1976a: 9–10). The ever, the wind measurements made at the Kevo seasonal ground frost remains in the open ar- Meteorological Station should be treated with eas generally until mid-June. caution because topography has a signiﬁ cant The mean annual precipitation was 395 effect on the wind pattern (see Mansikkaniemi mm and the average number of rainy days in & Laitinen 1990; Seppälä 2002). a year was 231 (63%). The mean cloud cover The climatological conditions of the was 5.9 octas (70–75%) and the mean relative Báišduattar – Áilegas area, especially above the humidity of air was ca. 80%. Precipitation was treeline, may differ considerably from circum-

Table 6. Summary of the climate parameters measured at the Kevo Meteorologi- cal Station (69°45’N, 27°01’E, 107 m a.s.l.). The values are from the period 1961– 1990 with some exceptions (for more details see text; Atlas of Finland 1987; Climatological Statistics in Finland 1961–1990 1991; Seppälä & Hassinen 1997).

Climate parameter Mean annual Air temperature -2.0°C Freeze / thaw index -2052.7 / 1372.7 Total temperature sum -681.6 Number of frost days (air / ground surface) 230 / 258 Frost depth (with / without snowcover) 130 / 210 cm Precipitation 395 mm Relative humidity 80% Number of snowcover days 210 Maximum snow thickness 69 cm Wind speed 2.9 m/s

31 Study area stances prevailing in the forested valley where 3.7 Hydrology the Kevo Meteorological Stations is located (e.g. Huovila 1987a; Seppälä & Hassinen 1997). The study area is on the watershed of the Fair differences could be expected to occur in Tenojoki (Deatnu) and Paatsjoki catchments. the temperatures, snow distribution and wind About 85% (518 km2) of the area is a part of speeds. The average wind speed has been the Tenojoki drainage basin (total area = 14 measured to be signifi cantly higher on the fells 891 km2 and percentage of lakes, L% = 3.1) (Mansikkaniemi & Laitinen 1990: 14–15). The and the rest of the area (82 km2) belongs to distribution of the snow is very dependent on the Paatsjoki catchment (14 512 km2, L% = the wind speed and direction, vegetation cov- 12.4) (Ekholm 1993). The main rivers are the er and topography (e.g. Mckay & Gray 1981: Deatnu (only 2 km in the study area), Áhkojoh- 154–166). For example, the snow cover can be ka and Čulloveijohka (Fig. 19). A number of in the birch forests from 50 cm to 70 cm, on smaller streams occur in the area and 20–25% the alpine heaths 20–30 cm (Kärenlampi 1972: of them are temporary, i.e. they dry during the 64), but on the fell summits snow thickness summer (Anonymous 2002). The largest lakes can be less than 5 cm and it can occasionally are Stuorrajávri (2.53 km2), Nuorttatjávri (1.94 disappear during the winter (e.g. Kallio et al. km2) and Biesjávrrit (0.57 km2) and the total 1969; Clark et al. 1985). In general, the snow cover of the lakes and ponds is ca. 8 km2 (L% thickness and duration have a signifi cant effect = 1.3) (Fig. 19). on the ground temperatures and thus on frost Hydrologically, the Báišduattar – Áilegas activity (e.g. Goodrich 1982). area is a typical subarctic region with a high The MAAT of the study area could be at but short-term fl ood peak (Fig. 23). The short maximum 3.45 degrees lower on the high- and drastic spring fl oods are the result of rapid est fells than at the Kevo if we only consid- snowmelt and the lack of signifi cant lake ba- er adiabatic lapse rate, but other factors can sins that would attenuate fl ood peaks. The sub- also obscure the temperature distribution (e.g. arctic nature of the Deatnu river is displayed Huovila 1987a; Holgrem & Tenow 1987). The by the high spring fl ood discharge, which is on air temperature is measured to be on the al- average ten to fi fteen times higher than winter- pine heaths from two to 3.5°C lower than in time discharge (Mansikkaniemi 1970: 6, 1972: the forested valleys in summer, but in winter, 15–16). In addition, the rise of the water level the fells can be generally 1–2 degrees warmer in the Deatnu river can be remarkable. Gen- due to inversion and cold air ponding in val- erally, the water rises during the spring fl oods leys (Kärenlampi 1972: 56–61; Helimäki 1974: from two to four metres over the normal level, 23–25). The difference can be over 20°C in but it may rise from fi ve to ten metres because midwinter but these extreme situations are of ice dams (Mansikkaniemi 1972: 16). rather temporary (Huovila 1987b: 24). Taking into account the previous facts, the MAAT of the study area is estimated to be between –2°C 3.8 Vegetation and –4°C depending on the altitude. However, it is worth of noting that considerable valleys The vegetation of the area is characterized by exist above the treeline in the northeast where subalpine mountain birch forests (Betula pubes- the cold air can drain during the inversions and cens ssp. czerepanovii) and alpine heaths (Fig. 19). thus the coldest locations do not have to be In the lower altitudes and in the deep river val- summit areas. leys occur small scattered scots pine (Pinus syl- vestris) forests. In general, forests cover about 37% of the study area (Anonymous 2002). In the mid-sixties, the larvae of the moth Epir- rita autumnata caused extensive birch forest damage in northernmost Finland (e.g. Kallio

32 Study area

Figure 23. Áhkojohka river valley after the spring ﬂ ood (69°36’27”N, 26°17’44”E / 400 m a.s.l. / 18th of June 2002). The water level has already lowered close to the mid-summer level. A snow patch on the right edge.

& Lehtonen 1973). Consequently, most of the m a.s.l. and the alpine vegetation occur above birches in the southeastern part of the study this elevation zone. area were defoliated (Fig. 24), although the The barren fell areas (regio alpina) are domi- cover of the damaged area was only ca. ten nated by Empetrum and Betula nana types of square kilometres (Seppälä & Rastas 1980). In alpine heaths (e.g. Fig. 21; Heikkinen & Kal- general, the recovery of the damaged areas has liola 1989). Total cover of the alpine heaths been rather slow (Fig. 24; e.g. Heikkinen & Ka- is about 270 km2 (ca. 45%), but this value also lliola 1989: 34). includes secondary alpine i.e. subalpine heaths Based on the vegetation zonation, the re- (for more details see Josefsson 1988; Heikki- gion belongs to the subarctic zone north of nen & Kalliola 1989: 27–28). On the fell sum- the northern limit of the continuous pine for- mits, the height of the vegetation is usually est (e.g. Hustich 1960) or to the orohemiarctic from a few centimetres to 20 cm and, in the zone (Ahti et al. 1968). In the continentality- valleys, the height can exceed 100 cm. The un- oceanicity division, the area is situated in the even distribution of the snowcover has a sig- continental sector (e.g. Oksanen & Virtanen niﬁ cant effect on the vegetation in the alpine 1995). The birch forests are mainly different belt (e.g. Kalliola 1939). Mires, which cover kinds of Empetrum types in the study region ca. 10% of the study area (Fig. 19), belong to (see e.g. Heikkinen & Kalliola 1989). The most the palsa and subalpine mire types (Ruuhijärvi common types are Empetrum-Lichenes and Em- 1960). In more detail, the most common types petrum-Lichenes-Pleurozium birch forests and the are heathy hummocky, Salix – Betula nana and former usually forms the altitudinal forest lim- dwarf shrub bog types (Heikkinen & Kalliola it. The treeline lies generally between 360–420 1989).

33 Study area

Figure 24. Only partly recovered open mountain birch (Betula pubescens ssp. czerepanovii) forest damaged by larvae of the moth Epirrita autumnata in the southeastern part of the study area (69°23’47”N, 26°27’1”E / 340 m a.s.l. / 5th of August 2003).

4 MATERIALS AND METHODS

4.1 Modelling data

4.1.1 Resolution

The modelling resolution (i.e. the size of the modelling square utilised in this study) was 25-ha (500 x 500 m). At fi rst, periglacial landforms were test modelled at four different resolutions (1, 4, 25 and 100 ha) (Hjort, unpublished data) but explicit performance differences did not occur when the models were compared using a critical ratio test (for more details see Pearce & Ferrier 2000). Therefore, Figure 25. The main steps of the compilation of the mesoscale resolution (cf. Summerfi eld response, i.e. periglacial, data. 1991: 13) was chosen based on: (1) the nature of the response variable, (2) the accuracy and quality of the predictor data and (3) the previ- topographic maps (1:20 000) were used in the ous modelling studies. Firstly, the distribution, preliminary survey. Many of the periglacial cover and accuracy of the response affect the landforms are relatively small-sized and their proper modelling resolution. For example, the direct identifi cation from the available aerial use of too fi ne resolution increases the sample photographs is impossible and, therefore, the size without a concomitant increase in prima- potential landform sites were mapped to be ry information while artifi cially enhancing the checked in the fi eld. Stereoscopic photo inter- degree of autocorrelation and pseudoreplica- pretation was conducted during the springs of tion in the statistical analysis (Hurlbert 1984). 2002 and 2003 and the results were drawn on Secondly, the quality of the databases from the working maps (1:10 000). which the predictors are compiled has to be Second, periglacial landforms were mapped taken into account (Chapter 4.1.3). Thirdly, a utilising pre-mapping results and black-and- 25-ha resolution has shown to be appropriate white aerial photographs (1:31 000) in the fi eld in periglacial and other landscape scale mod- during the summers of 2002 (18th of June–24th elling studies (Luoto et al. 2001, 2002; Luoto of October) and 2003 (10th of June–26th of Oc- & Hjort 2004, in press; see also Niessen et al. tober) in 100 fi eld workdays. The positions of 1992). the features were mapped with a GPS-device (Garmin eTrex personal navigator). Landform type and activity (active/inactive) was defi ned 4.1.2 Periglacial landforms visually in situ and the interpretations were documented in a notebook. The size of the Periglacial landforms were mapped and con- mapped landform was estimated and classifi ed verted to grid-based modelling data in a fi ve into four classes: <10 m, 10–30 m, 30–50 m or step process (Fig. 25). First, detailed stereo- > 50 m in diameter. The features less than 50 scopic interpretation of black-and-white aerial m were located with a single GPS point and, photographs (1:31 000) was performed to map if necessary, two or several points were used and identify periglacial features from the study to defi ne the borders of the larger landforms. area. In addition to the aerial photographs, If the features size could not be estimated, for

35 Materials and methods example due to scattered distribution, only the Attribute information was attached to each location was taken with the GPS-device. Fur- mapped landform. The fi nal database of the thermore, the size measurements with a tape periglacial landforms contained: (1) identifi ca- measure, and slope angle measurements with a tion number (id) for each object, (2) landform Suunto clinometer, were performed for differ- type classifi cation (indicated by code), (3) ac- ent types of small-sized landforms. tivity classifi cation (active/inactive) and (4) The activity of the features was defi ned size information (fi ve classes). based on the observations of lichen cover on Fourth, the maps of the active landforms the stones and blocks (Figs. 26A & B), rock were converted into 10 m raster grids in weathering, frost heaving, general soil and ArcView GIS 3.2 to calculate the cover of the vegetation disturbance (Fig. 26C) and vegeta- features in each modelling square. The land- tion density (Fig. 26D) (Goldthwait 1976: 34; form cover was calculated by the ZONAL Washburn 1979: 133; Cook-Talbot 1991: 128– function in the GRID module of Arc/Info 8.2 130; Harris 1994: 187). Moreover, previous and the results (in hectares) were used to pro- studies and activity measurements conducted duce the modelling data sets. The cover infor- in the region were used to gain an overall pic- mation was directly used in abundance model- ture of the activity of a specifi c landform type ling, if the modelling criteria were fulfi lled (see (see Chapter 3.5). If the mapped landform had below). A binary variable (1 = present, 0 = some indicator of activity, even on a relatively absent), indicating the occurrence of the land- small area, it was classifi ed as active (Luoto form, was allocated in each modelling square. & Hjort 2005; Hjort & Luoto 2006). Also, it The square was selected to be ‘present’ if the should be noted that a feature does not have cover was over zero hectares. In addition, the to be recently formed to be determined active features without size information were con- (Jahn & Siedlecki 1982; Ballantyne 1984). verted to binary form. Third, fi eld-mapping results were digitized Finally, landform types and modelling in a vector format on ortho-rectifi ed aerial squares were selected for the distribution and photographs utilising MapInfo Professional abundance modelling. The landform should 7.0 software. The ground resolution of the be fairly abundant to be used in statistical anal- digital aerial photographs was one metre. The yses (e.g. Guisan & Zimmermann 2000). How- location error of the ortho-rectifi ed aerial pho- ever, the proportion of the present squares tographs is estimated to be from three to six prevalence can vary considerably in binary metres in northern Finnish Lapland (National data (e.g. McPherson et al. 2004; Segurado & Land Survey of Finland 2004a) but the relative Araújo 2004). In this study, the prevalence had error between different images was signifi cant- to be 5–95%. This quite wide range was used ly more when planimetric root-mean-square because the other evaluation measure, namely error (RMSE) was calculated with 21 test sites. AUC of ROC plot (see Chapter 4.2.3), is prac- The RMSE was 18.0 m and it was calculated tically immune to prevalence-related artefacts with the equation: (McPherson et al. 2004: 822). The abundance modelling was focused on the present squares (1) and landforms types whose cover varied notably, respectively. where ei is the distance difference between In general, water cover can hamper the de- test points in metres (e.g. Zhang & Goodchild tection of real causal relationships between re- 2002: 76). This inaccuracy was taken into ac- sponse and predictor variables. To remove this count in the selection of the modelling reso- bias all squares containing water cover should lution (Chapter 4.1.1). Landform fi elds over be removed but this would reduce the number 50 m in diameter were vectorized as polygons of modelling squares too drastically. Thus, it in true spatial size and shape and features was decided to remove part of the water-cov- less than 50 m were digitized only as points. ered squares based on the data exploration.

36 Materials and methods

Figure 26. Active (A) (405 m a.s.l. / 16th of July 2003) and inactive (B) (400 m a.s.l. / 17th of July 2003) stony surface of sorted solifl uction sheet. The soil movement (solifl uction) has bent some of the mountain birches on a non-sorted solifl uction terrace (C) (69°27’11”N, 26°17’5”E / 330 m a.s.l. / 28th of June 2003). Inactive ca. 20 cm high earth hummocks covered by mosses (D) (69°27’27”N, 26°16’35”E / 345 m a.s.l. / 9th of July 2004).

In the distribution modelling, the land area and solar radiation, soil type and vegetation. should be over one hectare. In the abundance In addition, a group of spatial variables were modelling, the limit was strict; the proportion compiled to use in hierarchical partitioning of the land cover in the square should be over (Chapter 4.2.3). In general, spatial variables 95% because even relatively low water cover can be gathered from fi ve main sources: (1) may cause problems in the analyses. fi eld work, (2) digital and paper maps, (3) remote sensing, (4) maps obtained from GIS- based modelling and (5) other digital databas- 4.1.3 Predictor variables es (Guisan & Zimmermann 2000: 156). Here, the predictors were collected mainly from the The compilation and selection of appropriate second, fourth and fi fth groups (Fig. 27). The predictor (i.e. explanatory) variables for the fi eldwork is usually too laborious an informa- statistical analyses can be a complicated and tion collection method because the predic- diffi cult task without a comprehensive proc- tors should be more easily obtained than the ess-environment understanding. There are response (Guisan & Zimmermann 2000). neither universal criteria nor widely accepted Remote sensing (RS) provides an opportunity guidelines and hence the study aims guide to collect spatially continuous information of the procedure. The predictors used in this environmental factors but RS data were not di- study were collected from several information rectly used in this study. However, the utilised sources and classifi ed into fi ve main groups, vegetation information was based on fi eld sur- namely topography, soil moisture, temperature vey and remotely sensed data (Sihvo 2001).

37 Materials and methods

Figure 27. The main interactions between environmental factors affecting periglacial phenomena on a regional scale. The environmental predictors and surrogates of factors were derived from different data sources (outermost boxes).

The most commonly used and very versa- iglacial studies (e.g. Etzelmüller et al. 2001). tile source for predictors in spatial modelling is The altitude and slope angle are key surrogates a digital elevation model (DEM) (e.g. Moore et of many different environmental factors like al. 1991; Franklin 1995; Etzelmüller et al. 2001; air temperature, snow distribution and poten- Evans 2004). Therefore, a raster DEM with 20 tial energy of an area (Table 7). m grid size was calculated by linear interpola- A slope map of the study area was pro- tion from digital contour lines using Arc/Info’s duced from the DEM by the SLOPE function TOPOGRID command (ESRI 1991; see Fig. in the GRID module (ESRI 1991; Fig 28A). 20). The utilised interpolation method is de- The slope angle for each grid cell is calculated signed for the creation of hydrologically cor- from a 3 x 3 neighbourhood using the aver- rect DEMs (Hutchinson 1993). The contour age maximum technique (e.g. Zevenbergen & lines (5 m vertical interval) are digitized from Thorne 1987). Mean altitude in metres and paper copies of the Finnish topographic maps mean slope angle in degrees were calculated (1:20 000) (National Land Survey of Finland directly from the DEM and the slope map by 2004b). The quality of the created DEM was the ZONALMEAN command (ESRI 1991). evaluated visually and by calculating RMSE Standard deviation of altitude and slope angle with the equation (1) from height points (n was calculated by the ZONALSTD command, = 50) not used in the interpolation (for more respectively. In general, the standard deviation details see ESRI 1991; Hutchinson & Gallant of a parameter provides a more stable statistic 2000: 38–39). The vertical RMSE (relative ac- than ranges and extremes (Evans 1972: 31– curacy) was 2.2 m (mean error = 1.6 m). 32). However, the relative altitude, also called In this study, nine pure topographical variables local relief (e.g. Mark 1975: 167), was calcu- were calculated from the DEM using an Arc/ lated from the DEM by the ZONALRANGE Info GRID (Table 7). Altitude and slope angle command and maximum slope angel was cal- are commonly used topographical parameters culated from the slope map by the ZONAL- in general geomorphological (e.g. Evans 1972, MAX command. 2004; Mark 1975; Moore et al. 1991) and per- Elevation–relief ratio is a hypsometric varia-

38 Materials and methods

Table 7. Pure topographical variables and their description (Moore et al. 1991; Florinsky 1998; Etzelmüller et al. 2001).

Variable Description Mean altitude (m) Temperature, snow distribution, potential energy, radiation intensity, windiness, cloudiness, vegetation zone, moisture distribution Relative altitude (m) Relief variability, potential energy, water fl ow, moisture distribution Standard deviation of altitude (m) Topographical roughness, variability of temperature, snow distribution etc. (see mean altitude) Mean slope angle (°) Potential energy, water fl ow, snow distribution, radiation, soil thickness Standard deviation of slope angle (°) Variability of potential energy, water fl ow etc. (see mean slope angle) Maximum slope angle (°) Maximal potential slope energy (see mean slope angle) Elevation-relief ratio Topographical setting in a grid, distribution of land area in relation to altitude (hypsometric variable) Concave topography (%) Water, snow and sediment accumulation, soil thickness Flat topography (%) Moisture and snow distribution ble that describes the topographical setting in an use total curvature to determine the convexity area. This ratio value is the best approximation or concavity of an area (Mark 1975: 170; Gal- of the most common hypsometric value, name- lant & Wilson 2000: 56). ly the hypsometric integral (Mark 1975: 173). Four surrogates for soil moisture were calcu- The elevation–relief ratio (E) was calculated lated (Table 8). Three of them were derived with the equation (Pike & Wilson 1971: 1079): from the pit-fi lled DEM and one, namely water cover, from the vectorized basic map data. (2) Arc/Info’s FILL command was used to fi ll up sinks and level peaks to ensure a continuous where Z is the elevation. The value of the el- drainage network. Sinks and peaks are either evation–relief ratio varies form zero to one. errors in DEMs due to the resolution of the The value close to 0.5 indicates normally dis- data or rounding of elevations to the near- tributed and values close to zero or one very est integer value (ESRI 1991) or they may be skewed elevation distribution in the modelling natural depressions like lake basins and kettle square (Fig. 29A). holes. Concave topography represents accumula- The topographical wetness index is a tion areas for moisture and soil material, but commonly used indirect soil moisture indi- also for cold air during wintertime ground cator in square-based spatial analyses (e.g. inversions (e.g. Geiger 1965: 393–403; Harris Moore et al. 1991; Franklin 1995; Walsh et al. S.A. 1982b). The proportion of concave and 1998). The topographical wetness index (ω) fl at (slope < 2º) topography was calculated as was calculated using the following formula: a percentage for each modelling square. The α total curvature (see ESRI 1991) and slope ω = ln (As / tan ), (3) maps were used to calculate the proportion of concave and fl at topography. In general, cur- where ln denotes the natural logarithm, As vature can be divided into two parts, profi le represents the upslope contributing area and α (i.e. vertical) and plan (i.e. horizontal or con- the slope angle (Beven & Kirkby 1979; Moore tour) curvature, but it is more convenient to et al. 1991; Fig 28B). The mean, standard devi-

39 Materials and methods

Figure 28. Examples of the compiled environmental predictors at 20 m resolution: (A) slope angle, (B) topographical wetness index, (C) modelled minimum air temperature, (D) direct clear-sky short-wave solar radiation, (E) peat soil and (F) shrub cover (for more details see text). ation and maximum of the topographical wet- The minimum air temperature layer was con- ness index were calculated by ZONAL func- structed using multiple linear regression, point tions (see above; Fig. 29B). The water cover measurements of air temperature (n = 34) and for each modelling square was calculated from DEM-derived variables (Table 10; Luoto & the gridded water layer by the ZONALSUM Seppälä, unpublished data; Fig. 28C; cf. Vir- command (ESRI 1991). tanen et al. 1998). The mean of minimum air Temperature and solar radiation variables uti- temperature (ºC) of year was calculated direct- lised in this study are presented in Table 9. ly from the constructed temperature model by

40 Materials and methods

Figure 29. Examples of the compiled environmental predictors at 25-ha resolution: (A) elevation-relief ratio, (B) mean of topographical wetness index, (C) relative solar radiation and (D) mean of shrub cover. The DEM-based variables (A-C) included one square without data (water covered square) and shrub cover variable (D) 24 squares without information (lack of data in biotope database). The vertical interval of contours is 20 m (© National Land Survey of Finland, Licence number 49/MYY/06). the ZONALMEAN command. nal and autumnal equinoxes, with a 60 minute The relative solar radiation was calculated interval at the 20 m grid resolution. The short- from direct clear-sky short-wave radiation lay- wave radiation maps were generated by the ers in four steps. First, the direct short-wave DEM and an Arc/Info Arc Macro Language radiation was calculated for four days, namely (AML) routine. The utilised macro (short- summer and winter solstices as well as for ver- wavc.aml) is based on the assumption that the

41 Materials and methods

Table 8. Surrogates for soil moisture and their description (Moore et al. 1991; Wilson & Gallant 2000).

Variable Description Mean wetness index Water distribution, depth of water table, potential soil moisture, cold air distribution during inversion, organic matter, silt and sand content Standard deviation of wetness index Variability of water distribution etc. (see mean wetness index) Maximum of wetness index Maximal soil moisture etc. (see mean wetness index) Sum of water cover (100 m2) Moisture and heat energy storage

Table 9. Temperature and solar radiation variables and their description (Moore et al. 1991; Kumar et al. 1997).

Variable Description Minimum temperature of year (ºC) Frost intensity and depth Relative radiation (%) Potential radiation, temperature, snow, moisture and vegetation distribution, rock weathering Aspect majority (º) Potential radiation, snow and vegetation distribution

Table 10. Variables, coefﬁ cients, standard errors (SE), t- and p-values of the ﬁ nal model for minimum air temperature (Luoto & Seppälä, unpublished data).

Variable Coeﬃ cient SE t p Intercept -36.776 1.771 -20.765 <0.001 Mean altitude 0.032 0.003 11.158 <0.001 Mean wetness index -0.403 0.114 -3.539 0.001 Mean slope angle 0.217 0.077 2.818 0.008 transmittance of a clear sky is 0.8, and it cal- the relative solar radiation for each modelling culates the attenuation of the beam radiation square (Fig. 29C). attributed to lower sun-altitude angles (Kumar Aspect, which can be thought of as the et al. 1997; Zimmermann 2003). The utilised slope direction, is a common DEM-derived AML routine takes into account overshadow- surrogate for solar radiation and air tempera- ing by high peaks, meaning that the routine ture (e.g. Evans 1972, 2004; Moore et al. 1991; detects pixels that are in the shadow of ad- Franklin 1995; Etzelmüller et al. 2001). An jacent higher terrain for a given sun position. aspect map of the study area was produced Second, the average of the direct solar radia- from the DEM by the ASPECT function in tion of the four days was calculated for each the GRID module. The output value was the 20 m grid (Fig. 28D). Third, the mean of the compass direction of the aspect. The aspect estimate of solar radiation was calculated for majority was calculated directly from the as- each modelling square utilising ZONAL func- pect layer by the ZONALMAJORITY com- tion as previously (see above). Fourth, the ob- mand (ESRI 1991). tained values of radiation were divided by the Different soil types have dissimilar proper- maximum value of the study area to calculate ties upon which periglacial processes act (e.g.

42 Materials and methods

Table 11. Soil type variables and their description (Washburn 1979: 11, 14–15; Goodrich 1982; Williams & Smith 1989: 89–90).

Variable Description Peat cover (100 m2) High water-holding capacity, non-conductor (dry peat), dense to moderate vegetation coverage Glacigenic deposit cover (100 m2) Frost susceptible, high water-holding capacity, different size of soil particles Sand and gravel cover (100 m2) Frost resistant, dry, sparse ground layer vegetation Rock terrain cover (100 m2) Thin or absent soil cover, rock material, heat conductor

Table 12. Vegetation variables and their description (Josefsson 1988; Williams & Smith 1989: 61–62, 71–74; Liston et al. 2002).

Variable Description Mean shrub cover (%) Snow thickness and distribution, soil moisture and temperature Standard deviation of shrub cover Variability of snow thickness etc. (see mean shrub cover) Mean canopy cover (%) Snow thickness and distribution, air temperature Standard deviation of canopy cover Variability of snow thickness etc. (see mean canopy cover)

Washburn 1979: 11–15, 67–68). For example, gravel. Rock terrain was obtained from the bi- peat and glacigenic soils are more susceptible otope data to represents areas where the surfi - to ice lens formation and intensive frost heave cial deposits are thin, generally less than 50 cm than sand and gravel (Williams & Smith 1989: thick (cf. Eeronheimo 1996: 19–20). 30–32). The cover of four soil types was calcu- Vegetation cover is an important indirect en- lated for each modelling square from a digital vironmental factor but it has a diverse effect soil map by the ZONALSUM command. The on the process activity (e.g. Washburn 1979: soil map was constructed based on a Quater- 17; Clark et al. 1985; Matthews et al. 1998). nary deposit map 1:400 000 (Kujansuu 1981), Four variables, mean and standard deviation biotope data (Anonymous 2002) and fi eld of shrub cover and canopy cover, were com- observations. Soil types were fi rst digitized as piled to represent differences in the vegetation vectors and then converted to 20 m grids (e.g. coverage (Table 12). The vegetation informa- Fig. 28E). tion was obtained from the biotope data and The utilised soil types were peat, glacigenic the basis of mapping can be found in Eeron- deposit (till), sand and gravel as well as rock heimo (1996: 11–12) and Sihvo (2002: 39–40). terrain (Table 11). Peaty areas were determined The shrub cover varied between 0% and 85% based on the biotope data (Eeronheimo 1996: and the canopy cover from 0% to 60% in the 26–27) although the Quaternary deposit map study area. The calculations were performed, was used as a base data (Kujansuu 1981). The as previously, in Arc/Info GRID from raster thickness of the peat layer is estimated to be (20 m) data layers (Figs. 28F & 29D). generally more than 10 cm but it can be dis- Spatial variables were compiled for the hier- continuous. Glacigenic till was digitized based archical partitioning analyses to detect the in- on the Quaternary deposit map and it includ- dependent effect of environmental predictors ed blocky surfi cial material (see Chapter 3.3). (Table 13; see Chapter 4.2.3). Autocovariates Glaciofl uvial, fl uvial and aeolian deposits were that describe patch-like autocorrelation (Leg- determined based on the Quaternary deposit endre 1993: 1659) were calculated for each map and fi eld observations and they were clas- landform type used in the modelling: (1) the au- sifi ed into the same group, namely sand and tocovariate for binary response is the number

43 Materials and methods

Table 13. Spatial variables and their description (Legendre 1993; Franklin 1998).

Variable Description Spatial autocovariate Spatial autocorrelation of a speciﬁ c periglacial landform x-coordinate (m) Geographical trend, west-east aspect y-coordinate (m) Geographical trend, south-north aspect of occupied neighbouring squares divided by see Van Houwelingen & Le Cessie 1990). The the total number of neighbouring squares and calibration datum was used to adjust the mod- (2) the autocovariate for continuous response el (Chapter 4.2.2), whereas the evaluation da- is the total landform cover of neighbouring tum was used to evaluate the quality of model squares divided by the total number of neigh- predictions (Chapter 4.2.3). The random ap- bouring squares (Augustin et al. 1996; Luoto et proach was used to ensure the similarity of al. 2001; Heikkinen et al. 2004). The number the model calibration and evaluation sets (see of the neighbouring squares varied from three Pearce & Ferrier 2000: 237). The randomisa- to eight. In addition to autocovariates, centred tion was performed in Microsoft® Excel 2002 geographical coordinates, x and y, of the mod- by the RAND function. The evaluation data elling square were compiled (see Franklin 1998: cannot be considered totally independent 737). These variables describe possible trend because it is taken from the same area as the structures of the responses (Legendre 1993: calibration data. Consequently, evaluation data 1662). The utilised horizontal coordinate sys- should rather be kept as a quasi-independent tem was the Finnish National Grid Coordinate data set (Guisan & Hofer 2003: 1235). System (Gauss-Krüger projection with central longitude 27º) (see Atlas of Finland 1984). 4.2 Statistical modelling

4.1.4 Predictor variable selection and data split 4.2.1 Statistical formulation

The total number of environmental variables Statistical formulation means the choice of (1) was high (24) and some of them were highly an optimal statistical approach with regards correlated, Spearman’s rank correlation coef- to the modelling context and (2) a suitable al-

fi cient (Rs) up to 0.975, which could produce gorithm for modelling a particular type of re- severe statistical problems and uncertainties in sponse variable and estimating the model co- the modelling (Sokal & Rohlf 1981: 459–460). effi cients (Guisan & Zimmermann 2000: 158). Therefore, the most highly correlated envi- Based on the study aims (Chapter 1), data ex- ronmental variables (Rs > 0.75) were excluded ploration (distributions of the variables) and from the distribution and abundance model- previous studies (e.g. Etzelmüller et al. 2001; ling. This rather liberal limit was chosen be- Luoto & Seppälä 2002a; Luoto & Hjort 2004, cause, fi rstly, the effect of the multicollinearity 2005), the generalized linear modelling (GLM) could be explored by the hierarchical parti- method was chosen to be the main statisti- tioning method (Chapter 4.2.3) and, secondly, cal approach. The other essential modelling it was desired to gain versatile collection of method was hierarchical partitioning (Chapter predictors. When two predictors were highly 4.2.3). correlated, the selection was based on the reli- An important statistical development of the ability and interpretability of the variable. last decades has been the advance in regression The fi nal modelling data were randomly di- analyses provided by generalized linear mod- vided into model calibrating (70%) and model els (GLMs) (e.g. Nelder & Wedderburn 1972; evaluation sets (30%) (split-sample approach; Guisan et al. 2002). GLM is more fl exible and

44 Materials and methods

better suited for analysing geomorphological Further, Luoto and Seppälä (2002a, 2002b, relationships than the linear least square (LS) 2003), Lewkowicz and Ednie (2004), Luoto et regression method that has implicit statistical al. (2004a) as well as Luoto and Hjort (2004, assumptions (e.g. Atkinson et al. 1998: 1185). 2005, in press) have incorporated GLMs into Technically GLMs are relatively close to linear periglacial studies. regressions and thus relatively easy to utilise A binary (i.e. dichotomous) response varia- (e.g. Guisan et al. 2002). ble can have only two possible values, generally GLMs are mathematical extensions of coded as one and zero, for example present/ linear models that do not force data into un- absent or active/inactive. Binary data are often natural scales; they allow for non-linearity modelled with logistic regression that belongs and non-constant variance (heteroscedastic- to the class of GLMs (Collet 2003: 58). In lo- ity) structures in the data (e.g. Guisan et al. gistic regression, the relationship is expressed 2002). GLMs have three components: (1) the as a probability surface and the expected error

response variables y1, y2, …, yn, which are as- structure is binomial. Logistic regression has sumed to share the same distribution from the been demonstrated to be a powerful modelling exponential family, (2) a set of parameters α tool in solving different distribution problems and β and predictor variables and (3) a mono- (e.g. Nicholls 1989; Pereira & Itami 1991; Brito tone link function g, which allows transfor- et al. 1999; Bledsoe & Watson 2001; Dai & Lee mation to linearity and the predictions to be 2002; Luoto & Seppälä 2002a; Lewkowicz & maintained within the range of coherent val- Ednie 2004). ues for the response variable (McCullagh & A simple logistic regression model is a nor- Nelder 1989: 27). In GLMs, the explanatory mal regression model where the response is

variables xk are combined to produce a linear logged odds. The odds express the likelihood predictor (LP), also termed as a linear compo- of an occurrence (i.e. phenomena present) nent (e.g. Dobson 2002: 35), which is related relative to the likelihood of a non-occurrence. to the expected value (μ) of the response vari- Logistic regression can be presented as: able through a link function g():

α logit (pi) = log[pi/(1−pi)] = LP, (5) g(μ) = LP = + β1*x1i + β2*x2i + … + βk*xki, (4) where pi is the probability of the event and LP α where is a constant (intercept), βk are regres- is the linear predictor (Collet 2003: 58). Binary

sion coeffi cients and xki are predictor vari- GLM uses a logit link function; this means that ables. the probability of obtaining a positive response In GLMs, the model is optimised through is an s-shaped function when the LP is a fi rst- deviance reduction that is comparable to LS order polynomial and a bell-shaped when the models variance reduction (e.g. Guisan & Zim- LP is second-order polynomial (e.g. Crawley mermann 2000: 166). However, the regression 1993: 266). To calculate the probabilities as a coeffi cients of the model cannot be estimated function of the independent variables and the with the ordinary least square method. Instead, coeffi cients, an inverse logistic transformation maximum likelihood techniques, where the es- has to be used (e.g. Collet 2003: 58): timation method maximises the log-likelihood p = exp(LP)/[1+exp(LP)]. (6) function, are used to calculate these parameters i (e.g. Collet 2003: 52). Fitting the GLM is much the same as fi tting the multiple LS regression This transformation is necessary to obtain where polynomial terms can be included in probability values between zero and one. More the set of predictors to account for non-linear detailed description of the logistic regression and multi-modal responses. More information method is presented by Cox and Snell (1989), about the GLMs can be found from McCul- McCullagh and Nelder (1989), Dobson (2002) lagh and Nelder (1989) and Dobson (2002). and Collet (2003).

45 Materials and methods

Moreover, univariate analyses were per- Crawley 1993: 192–193; Venables & Ripley formed to explore, separately, the roles of each 2002: 175–176). The selection started with a environmental variable on the distribution and full model where all predictors were fi tted and abundance of the periglacial landforms. Sev- at each step one model component was manu- eral means to conduct univariate analysis exist ally omitted. Elimination was based on strict (e.g. Pereira & Itami 1991; Luoto et al. 2001). criterion (p < 0.001) for variable exclusion. Here, all environmental variables and their This very low p-value was chosen because of quadratic terms were included into the model the relatively high sample size, which tends to and in each step one explanatory variable at a lower p-values compared to smaller sample time was omitted and the change of the resid- sizes (McBride et al. 1993: 425). After all the ual deviance was calculated (residual deviance non-signifi cant predictors were eliminated, of the whole model minus residual deviance four to six omitted but potentially important after the exclusion) (see Crawley 1993: 192). predictors were tested for reinclusion into the model. The fi nal GLM model was fi tted with the selected signifi cant terms only. 4.2.2 Model calibration The proportion of the explained deviance [(null deviance – residual deviance / null devi- In model calibration, (1) the environmental ance) * 100] was calculated for the fi nal distri- variables will be selected to the fi nal model bution and abundance models to gain an overall and (2) the statistical model will be construct- picture of the success of the fi tting. In addition, ed (e.g. estimation and adjustment of model residuals were used to assess the success of the parameters) (Guisan & Zimmermann 2000: calibrations and the appropriateness of the se- 166). The model calibration was performed lected probability distributions used for the re- using the statistical package R version 1.8.0, sponses in abundance modelling (e.g. Crawley with standard glm function (Dalgaard 2002: 1993: 211–212; Dobson 2002: 19). 192; Venables & Ripley 2002: 183–210). The possibility of curvilinear relationships between explanatory and dependent variables were ex- 4.2.3 Model evaluation amined by including the quadratic terms of the predictors in the models (e.g. Crawley 1993: Evaluation of the generated model is a vital 192; Atkinson et al. 1998: 1185). The calibra- step in the model building process (e.g. Field- tion was started by choosing link functions for ing & Bell 1997; Guisan & Zimmermann the GLMs. A logit link function was chosen 2000: 171–174; Pearce & Ferrier 2000; Rush- for the binary responses (McCullagh & Nelder ton et al. 2004: 198). In this study, models were 1989: 31–32). A Box-Cox transformation (yλ) evaluated (1) quantitatively utilising evaluation approach was used to defi ne the optimal trans- data sets (discrimination measures and corre- formation to normalise the distribution of the lations), (2) descriptively and empirically using continuous responses (Box & Cox 1964; cf. prediction maps and fi eld observations (e.g. Dobson 2002: 109; Oksanen 2003: 66–67). Guisan & Zimmermann 2000: 172) and (3) the The Box-Cox function was performed using results (environmental variables) of the GLM the ‘MASS package’ and ‘boxcox’ command models were compared with the hierarchical in the statistical software R 1.8.0 (Venables & partitioning results (see below). The distribu- Ripley 2002: 170–172). An identity link was tion models were evaluated with all three ap- then used for the normalised response (Mc- proaches and abundance models with the fi rst Cullagh & Nelder 1989: 31–32). and third method. The variables were selected using a statisti- The prediction ability of the abundance cally focused backward elimination approach, models were evaluated quantitatively by calcu- in which predictors are excluded simply ac- lating Spearman’s rank correlation coeffi cient cording to their statistical signifi cance (see (Rs) between the predicted and observed val-

46 Materials and methods ues (e.g. Guisan & Zimmermann 2000: 173). on the x-axis. This makes the area under the There are not generally accepted measures of ROC function a threshold-independent and performance for binary models, respectively an unbiased measure (e.g. Pearce & Ferrier (e.g. Fielding & Bell 1997). Therefore, the dis- 2000: 232). crimination ability of the distribution models The AUC ranges from 0.5 for models with were assessed with two different measures: (1) no discrimination ability to one for models Cohen’s Kappa statistic (κ) (Cohen 1960) and with perfect discrimination. For example, a (2) the AUC value (Metz 1978). Both utilised value of 0.9 for AUC means that the model discrimination measures were calculated by can correctly discriminate between an occu- the SPSS for Windows 11.0 software package. pied and unoccupied site 90% of the time. In The most commonly used measure, namely other words, if a pair of evaluation sites (one the percentage of correct classifi cation (PCC), where phenomena is present and the other was not utilised because PCC can be mislead- where it is absent) is chosen at random, then ingly high when frequencies of zeros and ones there is a 0.9 probability that the model will in binary data are very different (Pearce & Fer- predict higher likelihood of occurrence for rier 2000: 230–231). the present site than for the absent site. Swets The Cohen’s Kappa coeffi cient measures (1988: 1292) proposed a rough guide for classi- the correct classifi cation rate, proportion of fying the AUC values of the models: 0.50–0.70 correctly classifi ed presences and absences, = poor, 0.71–0.90 good, > 0.90 = excellent. after the probability of chance agreement has In addition to the different evaluation sta- been removed (Cohen 1960; Congalton 1991). tistics, the prediction ability of the distribution According to Landis and Koch (1977: 165) models was evaluated visually and empirically models can be classifi ed based on the Kappa utilising prediction maps and new fi eld obser- statistics to: poor κ < 0.4; good 0.4 < κ 0.75 vations. A total of 136 modelling squares that and excellent κ > 0.75. The Kappa value is were mapped as unoccupied but had a relatively dependent on a single threshold to distinguish high predicted probability of landforms pres- between predicted presence and absence and ence [cf. commission errors (e.g. Congalton thus falls into the class of threshold–depend- 1991)] were visited in the fi eld during the sum- ent measures. An optimum probability thresh- mer 2004 (8th of July–21st of July). The causes old (accuracy of 0.01) was explored for each of low probabilities of the present sites [cf. distribution model based on the κ values of omission errors (e.g. Congalton 1991)] were the calibration data (e.g. Segurado & Araújo detected using extrapolated prediction maps 2004: 1559). and obtained fi eld knowledge. The predictions In general, a problem with the threshold- for the whole study area were calculated using dependent indices is their failure to use all of calibration models. the information provided by the classifi er. One Multivariate models are subjected to two threshold-independent measure has received notable problems: (1) predictor variables are more attention in medical and ecological lit- frequently signifi cantly intercorrelated (multi- erature (e.g. Zweig & Campbell 1993; Fielding collinearity), which can produce spurious cau- & Bell 1997: 44–46; Pearce & Ferrier 2000: sality; and (2) specifi city of the derived models 232–237), but also more recently in perigla- to the used data set that makes model extrapo- cial studies (Luoto et al. 2004a; Luoto & Hjort lation questionable (MacNally 1996: 228). The 2005; Hjort & Luoto 2005, 2006), than others; hierarchical partitioning (HP) method is de- the area under the curve (AUC) of a receiver signed to overcome these problems by using operating characteristic (ROC) plot. A ROC mathematical hierarchical theorem by which plot is obtained by plotting sensitivity values the explanatory capacities of a set of inde- (true positive proportion) on the y-axis against pendent variables can be estimated (Chevan & their equivalent (1 – specifi city) values (false Sutherland 1991: 92–94). HP does not produce positive proportion) for all possible thresholds a regression model but it helps to make better

47 Materials and methods deductions of the important environmental (1996, 2000, 2002), MacNally and Horrocks determinants. Furthermore, if two modelling (2002) and Heikkinen et al. (2004). For more approaches, in this study GLM and HP, agree detailed description of the method see Chevan on which predictors are the most important, and Sutherland (1991) and MacNally (1996, then it is more likely that meaningful predic- 2002). tor variables have been found (e.g. MacNally In this study, hierarchical partitioning was 2002: 1397). used to reveal the most likely causal variables HP employs goodness-of-fi t measures for by removing the effect of spatial autocorrela- each of the 2k possible models for k independ- tion as well as geographical trend by introduc- ent variables. An appropriate goodness-of-fi t ing spatial variables into the partition routine. measure depends on the distribution of the re- HP was conducted using ‘hier.part package’ sponse variable (MacNally 1996: 225). In HP, version 0.5–1 in statistical software R version the variances are partitioned so that the total 1.8.0 (Walsh & MacNally 2003). The maxi- independent contribution (I) of a given predic- mum number of independent variables that tor variable is estimated. Furthermore, the var- can be used in the partition is 12. This limited iation shared with another variable (i.e. cojoint the number of environmental variables to nine contribution, J) can be computed (Chevan & because three spatial variables (Chapter 4.1.3; Sutherland 1991: 92–93). For example, the in- Table 13) were included in every HP run. The dependent impact of variable ‘mean altitude’ is environmental variables were chosen based on estimated by comparing goodness-of-fi t meas- GLM modelling so that all variables in the fi - ures for all possible models involving ‘mean nal model were included. Additionally, if these altitude’ to yield the independent explanatory variables and spatial variables did not sum up power of this variable. In other words, the in- to 12, potentially important variables excluded crease in model fi t generated by ‘mean altitude’ from the fi nal GLM model were also included. is estimated by averaging its infl uence over all Used goodness-of-fi t measures were log-like- models in which ‘mean altitude’ appears. By lihood for binary data and root-mean-square these means, HP allows one to identify those prediction error (RMSPE) for abundance data predictor variables whose independent, as dis- (Walsh & MacNally 2003). For every response tinct from partial, correlation with a response variable there was generated 4096 regression variable may be important from variables that models to fi nd out the most meaningful pre- have little independent effect on the studied dictors. In the end, the interest was focused phenomena. The HP approach has been used, on the independent contribution of specifi c for example, in ecological studies by MacNally predictor variable.

5 RESULTS

5.1 Predictor variables were classifi ed into six main groups (Table 15). Palsas and thermokarst ponds are the only fea- In this study, a large number of predictors that tures in the fi rst group. Patterned ground is the are potentially important in modelling perigla- largest group with six main landform classes, cial phenomena were calculated for each mod- namely circles, polygons, nets, boulder depres- elling square at the 25-ha resolution. The fi nal sions, steps and stripes. Circles have seven sub- array of predictors utilised in the modelling classes of which four represented non-sorted included 15 environmental factors and all fi ve and three sorted features. In general, non- variable groups were represented (Table 14; see sorted circles are clearly more abundant than also Chapter 4.1.4). Mean altitude, mean slope sorted features forming extensive fi elds in the angle, elevation-relief ratio and proportion of central parts of the study area (Fig. 30). Sorted concave topography were selected from the circles, however, have rather scattered distribu- group of pure topographical predictors. The tion, although their occurrence is connected to maximum of topographical wetness index the non-sorted circles and sorted nets. was omitted from the soil moisture variables, Polygons and stripes include both non- whereas mean and standard deviation of wet- sorted and sorted landforms, whereas nets are ness index and water cover were taken into the all sorted and steps non-sorted. In the general- analyses. The mean of minimum air tempera- ised periglacial geomorphological map, sorted ture was excluded and relative solar radiation polygons are included in the class of sorted as well as aspect majority were selected from nets because polygons occur primarily in the the climatological factors. Mean of shrub cov- same topographical locations as nets, mainly er and mean of canopy cover were taken from on the fell summits. Boulder depressions are the group of vegetation variables. Moreover, common on the edges of mires, close to the all soil variables were included into the fi nal occurrences of non-sorted and sorted circles. set of predictors. Sorted stripes are closely connected to the Some of the selected variables were still sorted solifl uction sheets and therefore, only rather highly correlated (Table 14). The highest some separate stripe occurrences are indicated bivariate correlations (Rs > 0.6) were between by symbols in Figure 30. Non-sorted stripes canopy cover and altitude (Rs = −0.729), peat are spatially connected to the non-sorted cir- and shrub cover (Rs = 0.696), glacigenic de- cles and steps. posit and peat (Rs = −0.690), concave topog- Slope phenomena include four landform raphy and elevation-relief ratio (Rs = −0.645) groups, of which solifl uction is the largest with as well as between mean of wetness index and eight feature types. Terraces, lobes, ploughing concave topography (Rs = 0.633). Relative blocks and braking blocks are non-sorted and radiation and aspect majority correlated least sheets, terraces, lobes and streams sorted solif- with other variables. luction landforms. Rapid slope phenomena and taluses have fairly limited distribution in the study area (Fig. 30). The distribution of 5.2 Periglacial landforms in non-sorted solifl uction landforms is scattered Báišduattar – Áilegas at the whole study area scale. In addition, the different feature types are not spatially con- Periglacial landforms are common geomor- nected to each other or with any other per- phological features in the study region and the iglacial landform types. The sorted solifl uction diversity of the phenomena is rather consider- landforms occur in similar regions, on fell able (Figs. 30 & 31). The mapped landforms slopes at high altitudes (Fig. 30).

Results

cances were derived from the bivari- the from derived were cances

cant)].

signiﬁ statistical and cients

< 0.05, n.s. = not signiﬁ not = n.s. 0.05, <

. Correlation . coefﬁ

< 0.01, * = = * 0.01, <

< 0.001, ** = = ** 0.001, <

*** = = ***

the environmental predictors used in the statistical analyses statistical the in used predictors environmental the

[Spearman’s rank correlation rank ( [Spearman’s

ate correlation ate procedure Table 14. Correlation 14. of matrix Table

50 Results

Block fi elds and tors are the only frost subsequently. A total of 2272 squares of 25-ha weathering features mapped from the study were used in the statistical analyses, whereas area. The block fi elds are present at fairly high 128 squares were excluded based on the mod- altitudes but not frequently on fell summits elling square selection (Chapter 4.1.2) and due where tors are abundant. The group of nival to the defi ciency in the biotope data. phenomena includes snow accumulation hollows and stone pavements, which are common in the bottoms of glaciofl uvial channels. The 5.2.1 Palsas last landform class, aeolian features, includes sand dunes and defl ation landforms. Dunes Palsas and palsa mires represent four dif- occur mainly in the southeastern corner east of ferent morphological types: (1) string-form the Luopmošjáguotkku esker where also defl a- palsas, (2) hummocky palsas, (3) conical (i.e. tions are common, although defl ations can be dome-shaped) palsas and (4) palsa complexes. found from different parts of the study area String-form palsas are generally one metre (Fig. 30). In addition to the features listed in high, 1–2 m wide and often 10–30 m long tor- Table 15, few lakeshore and fl uvial landforms tuous ridges but features can be several tens were observed but they are not treated here of metres long (Fig. 32). Hummocky palsas because their periglacial nature is unclear. are circular or longitudinal from one to 1.5 In the Báišduattar – Áilegas area, the most m high and up to three metres in diameter. common periglacial landforms, if calculated by Dome-shaped palsas are from one to four me- occupied modelling squares, were earth hum- tres high and generally some tens of metres in mocks (n = 1636) and peat pounus (n = 702), diameter (Fig. 33). Most of the dome-shaped sorted nets (n = 631), sorted sheets (n = 431) palsas have fi ssures at their surfaces and some and defl ations (n = 295). The rarest features of the steepest edges have collapsed due to were slushfl ow tracks (n = 4), braking blocks block erosion (Fig. 33). Dome-shaped palsas (n = 4), non-sorted lobes (n = 10), plough- and thermokarst ponds (Fig. 4) are frequently ing blocks (n = 15) and stone pavements (n = found in palsa complexes (Fig. 2). Palsa com- 15), respectively. Snow accumulation hollows plexes have different morphological types of were mapped only during the summer 2002 palsas at different development stages. How- because their interpretation from the land- ever, only few examples of recently formed scape was shown to be rather diffi cult and the palsas were detected from the study area. All results unreliable. Therefore, the total number palsas were determined to be active. of the snow accumulation features was not The whole distribution of palsas can be seen determined. in Appendix 1. The largest palsa areas, from ca. The number of different periglacial land- six hectares to 21 ha, are found from the mires form types in the 25-ha squares varied from of Biesjeaggi, south of Girjeeana, Bihtoš-Per zero to nine (Fig. 31). A total of 247 (9.9%) jeaggi, Luopmošjohjeaggi and Vanađanjeaggi. squares were without any periglacial features The total cover of the palsas is about 227 ha and 2253 (90.1%) sites were occupied by at in the study area. Topographical characteristics least one landform type. Two types were ob- of present palsa squares in relation to absent served from 671 squares and this class was sites are presented in Table 16 and Figure 34A. the largest. Over six landform types were ob- The mean altitude and mean slope angle dif- served only from 58 (2.3%) squares. fered signifi cantly between the two groups (p Hereafter the focus will be on the perigla- < 0.001 in Mann-Whitney U-test). cial landforms utilised in the statistical analyses (see Chapter 4.1.2). A general morphological and distributional description of the modelled features as well as topographical characteristics of the active landform squares is provided

51 Results

52 Results

Figure 30. Generalised geomorphological map of the periglacial phenomena observed in the study area. Sorted net areas include sorted polygons and sorted solifl uction sheet fi elds enclose sorted stripes and solifl uction streams. The vertical interval of contours is 20 m (© National Land Survey of Finland, Licence number 49/MYY/06).

5.2.2 Convex non-sorted circles The largest continuous areas are situated in the valleys near the fells of Guivi, Leakšagoađoaivi Convex non-sorted circles are from one to and Suophášoaivi. The sizes of the mapped four metres in diameter and ca. fi ve to 60 cm feature areas range from ca. 100 m2 to 30 ha. high convex or fl at-centered mineral soil hum- The total cover of the active convex non- mocks (Figs. 6 & 35). The shape of the circle sorted circles is about 320 ha. Topographical can be elongated or step-like on gentle slopes, characteristics of active landform squares are and they may transform to non-sorted steps presented in Figure 34B and Table 16. The on slopes of more than three degrees (cf. mean altitude and mean slope angle differed Chapter 2.2.3). The surface of the circle can signifi cantly between the occupied and unoc- be characterised by small non-sorted or sorted cupied squares (p < 0.001 in Mann-Whitney polygons (Fig. 36). The inter-circle sections U-test). are usually covered by 10–40 cm high vegetation but some are moderately sorted (Fig. 35). Hence, convex non-sorted circles can evolve 5.2.3 Stony earth circles to sorted patterned features (sorted circles or nets) on level ground if the sorting proceeds. Stony earth circles’ diameters range from ca. Almost all circles were unvegetated and thus 0.1 to 1.5 m (Fig. 37). The shape of the circle active. Inactive features were found only from varies from circular to elongate. On fairly steep a few sites (Table 15). slopes, circles may transform to non-sorted Convex non-sorted circles are common in stripes (Chapter 2.2.3). If two circles coalesce, the northeast part of the study area where they they can form sandglass-like features. Only ac- form extensive pattern fi elds (Appendix 2). tive unvegetated bare-ground circles were de-

53 Results

Figure 31. Number of different periglacial landform types in the study area at 25-ha resolution. The vertical interval of contours is 20 m (© National Land Survey of Finland, Licence number 49/MYY/06).

54 Results

Table 15. Periglacial landforms and their prevalence in the Báišduattar – Áilegas area at the 25-ha modelling resolution (- = not observed, x = observed but not determined). Main division Periglacial landforms Present squares, Present squares, Present squares, active (%) inactive (%) total (%) Permafrost and Palsas (4 subtypes, see thermokarst Chapter 5.2.1) 134 (5.4%) - 134 (5.4%) Thermokarst ponds 38 (1.5%) 63 (2.5%) 69 (2.8%) Patterned ground Circles convex non-sorted circles 221 (8.8%) 14 (0.6%) 229 (9.2%) stony earth circles 245 (9.8%) - 245 (9.8%) earth hummocks 1609 (64.4%) 238 (9.5%) 1636 (65.4%) peat pounus 692 (27.7%) 25 (1.0%) 702 (28.1%) sorted stone circles 62 (2.5%) 21 (0.8%) 81 (3.2%) stone pits 247 (9.9%) 5 (0.2%) 252 (10.1%) debris islands 28 (1.1%) 11 (0.4%) 38 (1.5%) Polygons non-sorted polygons 9 (0.4%) 9 (0.4%) 18 (0.7%) sorted polygons 3 (0.1%) 48 (1.9%) 50 (2.0%) Nets sorted nets 188 (7.5%) 499 (20.0%) 631 (25.2%) Boulder depressions 9 (0.4%) 68 (2.7%) 73 (2.9%) Steps non-sorted steps 111 (4.4%) - 111 (4.4%) Stripes non-sorted stripes 35 (1.4%) 1 (0.0%) 36 (1.4%) sorted stripes 186 (7.4%) 15 (0.6%) 198 (7.9%) Slope phenomena Debris fl ow slopes 17 (0.7%) - 17 (0.7%) Slushfl ow tracks xx4 (0.2%) Solifl uction features non-sorted terraces 267 (10.7%) 10 (0.4%) 275 (11.0%) non-sorted lobes 9 (0.4%) 1 (0.0%) 10 (0.4%) ploughing blocks 9 (0.4%) 6 (0.2%) 15 (0.6%) braking blocks 2 (0.1%) 2 (0.1%) 4 (0.2%) sorted sheets (2 subtypes, see Chapter 5.2.10) 418 (16.7%) 39 (1.6%) 431 (17.2%) sorted terraces xx74 (3.0%) sorted lobes xx68 (2.7%) sorted streams 236 (9.4%) 14 (0.6%) 244 (9.8%) Talus slopes 12 (0.5%) 7 (0.3%) 19 (0.8%) Frost weathering Block fi elds - 34 (1.4%) 34 (1.4%) Tors xx127 (5.1%) Nival phenomena Snow accumulation hollows xx x Stone pavements xx15 (0.6%) Aeolian phenomena Sand dunes (parabolic, transverse and longitudinal) - 99 (4.0%) 99 (4.0%) Defl ations (2 subtypes, see Chapter 5.2.12) 295 (11.8%) 295 (11.8%)

55 Results

Figure 32. String-form palsa in Luopmošjohjeaggi mire (69°23’7”N, 26°10’26”E / 290 m a.s.l. / 22nd of August 2003). The ridge is ca. one metre high, two metres wide and some tens of metres long.

Figure 33. Ca. 2–3 m high dome-shaped palsas in Vanađanjeaggi mire (69°33’58”N, 26°15’6”E / 365 m a.s.l. / 9th of July 2002). Note the block erosion on the sides of these mature palsas. tected from the study area (Table 15). on the Mann-Whitney U-test, the altitude (p The distribution of stony earth circles is = 0.160) and slope angle (p = 0.476) of the relatively dispersed in the study area but some present and absent sites did not differ signiﬁ - clusters can be found from southeast, around cantly. Áilegas fell and in the wide zone east of Uhc- Áhkovárás fell (Appendix 3). Only the location of the features was mapped because of 5.2.4 Earth hummocks their scattered distribution. Topographical characteristics of stony earth circle squares are Earth hummocks’ heights range from ca. 0.1 presented in Figure 34C and Table 16. Based to 0.9 m and diameter from 0.2 m up to 1.5 m

56 Results

Table 16. Topographical characteristics of the active landform occurrences in relation to absent sites. The analyses were performed with 2272 25-ha squares and 128 squares were excluded. Statistical signiﬁ cances (p) of the differences between present and absent squares were derived from the Mann-Whitney U-test (*** = p < 0.001, ** p < 0.01, * p < 0.05, n.s. = not signiﬁ cant) and are illustrated in the present square columns only.

Periglacial landform Present Absent Altitude (m a.s.l.) Slope angle (°) Altitude (m a.s.l.) Slope angle (°) mean ± std mean ± std mean ± std mean ± std Palsa 338 ± 32*** 1.6 ± 1.0*** 383 ± 58 4.8 ± 2.4 Convex non-sorted circle 421 ± 42*** 3.5 ± 1.5*** 376 ± 57 4.7 ± 2.5 Stony earth circle 382 ± 41 n.s. 4.4 ± 2.1 n.s. 380 ± 59 4.6 ± 2.5 Earth hummock 361 ± 45*** 4.0 ± 2.2*** 422 ± 60 5.9 ± 2.6 Peat pounu 348 ± 38*** 3.0 ± 1.7*** 394 ± 59 5.2 ± 2.5 Stone pit 354 ± 45*** 2.9 ± 1.7*** 384 ± 58 4.8 ± 2.5 Sorted net 390 ± 53** 3.4 ± 1.8*** 380 ± 58 4.7 ± 2.5 Sorted stripe 467 ± 54*** 7.3 ± 2.2*** 373 ± 51 4.3 ± 2.4 Non-sorted terrace 407 ± 55*** 5.3 ± 2.3*** 377 ± 57 4.5 ± 2.5 Sorted sheet 447 ± 60*** 7.3 ± 2.5*** 366 ± 46 4.0 ± 2.1 Sorted stream 462 ± 55*** 7.6 ± 2.3*** 371 ± 50 4.2 ± 2.3 Deﬂ ation 371 ± 42 n.s. 4.0 ± 1.9** 382 ± 60 4.7 ± 2.5

(Fig. 38). Small features are often silty mineral vegetation cover and they can have cracked or soil, whereas large hummocks have peat cover mudboil-like summits, whereas mosses cover on mineral soil core. The surface of the hum- inactive hummocks (Fig. 26D). mock is usually characterised by vegetation Earth hummocks are prevalent periglacial cover. Cryoturbated internal structures were landforms in the Báišduattar – Áilegas area found in excavations performed in the valleys (Appendix 4; Table 15). The largest continuous of Leakšagoahti and northwest of Ruohtir fell. feature areas are located in Girjeeana region, Active earth hummocks have relatively sparse east of Uhc-Áhkovárás fell, south of Gaska-

Figure 34. Mean altitude and mean slope angle of the present and absent (A) palsa, (B) convex non-sorted circle and (C) stony earth circle squares. Only active landform occurrences were used in the analyses.

57 Results

Figure 35. Convex non-sorted circles on a gentle slope west of Ráššoaivi fell (69°38’29”N, 26°16’23”E / 495 m a.s.l. / 12th of August 2002). The circles are from 10 cm to 30 cm high and their diameter ranges from one to three metres.

Figure 36. Convex non-sorted circle with small partly vegetated (mainly Calluna vulgaris) non-sorted polygons on the surface (69°30’37”N, 26°28’57”E / 410 m a.s.l. / 30th of July 2002). The circle is ca. two metres in diameter and the cracks are from ﬁ ve to 20 cm wide.

Biesvárri fell, around Viercagálmaras fell as earth hummock squares in relation to absent well as southwest and southeast of Geatgie- sites are presented in Figure 39A and Table 16. las fell. The sizes of the mapped areas range The mean altitude and mean slope angle dif- from ca. 100 m2 to several square kilometres. fered signiﬁ cantly between the two groups (p The total cover of earth hummocks is ca. 70 < 0.001 in Mann-Whitney U-test). km2. Topographical characteristics of active

58 Results

Figure 37. Typical stony earth circle (69°28’0”N, 26°15’50”E / 395 m a.s.l. / 25th of June 2003). The diameter of the feature is ca. 80 cm.

Figure 38. Ca. 20–40 cm high active earth hummocks (69°34’10”N, 26°14’55”E / 370 m a.s.l. / 9th of July 2002). Dwarf birches (Betula nana) are common in the hummock ﬁ elds.

5.2.5 Peat pounus located in thin peat deposits or on mire edges. Most of the features were active (Table 15). Peat pounus are from ca. 0.2 m to 1.2 m high Inactive pounus are characterised by dry peat vegetated hummocks (Fig. 40). Pounus diam- and are often partly destroyed. eter ranges from ca. 0.3 m up to two metres The general distribution of pounus is pre- and their shape is mostly circular but features sented in Appendix 5. The largest feature can be elongated. The largest pounus are often ﬁ elds are found from Čulloveijeaggi mire,

59 Results

Figure 39. Mean altitude and mean slope angle of the present and absent (A) earth hummock, (B) peat pounu and (C) stone pit squares. Only active landform occurrences were used in the analyses.

Biesjeaggi mire, south and southwest of Gir- 5.2.6 Stone pits jeeana, southwest of Njávgoaivi fell and from Luopmošjohjeaggi mire. The sizes of the Stone pits’ diameters range from 0.3 to two mapped feature areas range from ca. 100 m2 to metres and their depth is generally less than 243 ha and the total cover of the peat pounus one metre (Fig. 41). The material of the fea- is 1931 ha. Topographical characteristics of tures is stone or block-sized (0.1–1 m). Stone active peat pounu occurrences are presented pits are more or less circular but if they coa- in Figure 39B and Table 16. The mean altitude lesce, the shape of the pit can be more com- and mean slope angle differed signiﬁ cantly be- plex. Most of the mapped features were active tween the present and absent sites (p < 0.001 and only few inactive features were detected in Mann-Whitney U-test). (Table 15). Inactive stone pits were covered by vegetation, usually mosses. Stone pits are fairly common and quite

Figure 40. Peat pounus in the western part of Čulloveijeaggi mire (69°35’21”N, 26°13’10”E / 365 m a.s.l. / 16th of July 2004). The hummocks are from 40 cm to 80 cm high.

60 Results

Figure 41. Active stone pit (69°37’35”N, 26°28’10”E / 490 m a.s.l. / 22th of June 2002). The depth of the pit is ca. 50 cm and the diameter is ca. one metre.

sporadically distributed despite some feature less than two degrees, the shape of the stony concentrations existing (Appendix 6). Stone mesh resemble polygonal but on gentle slopes pits can form relative continuous fi elds where (2–6º) it is clearly elongated. Most of the the distances between the features are 1–5 sorted nets are moderately or poorly sorted. m. However, more often their distribution is However, optimal conditions can create well quite random and the distances vary from a sorted features. The active nets are partly or few metres to several tens of metres. Due to totally without vegetation cover, whereas inac- the scattered distribution, only the location of tive features have lichen on stones and dense the features was documented. Topographi- vegetation cover in the centres (Fig. 43). cal characteristics of active present stone pit Sorted nets can be found from different squares are presented in Figure 39C and Table parts of the study area (Appendix 7). The 16. The mean altitude and mean slope angle largest areas, which are inactive, are in the fell differed signifi cantly between the present and regions of Áilegas, Njávgoaivi – Suohpášoaivi absent squares (p < 0.001 in Mann-Whitney and Guovdoaivi – Gamoaivi. The largest ac- U-test). tive feature areas were in Leakšagoahti valley, northwest of Biesjeaggi mire, south of Gamo- aivi fell and northwest of Leakšagoađoaivi 5.2.7 Sorted nets fell, respectively. In general, the sizes of the mapped fi elds ranged from ca. 100 m2 to over The size of the sorted nets (i.e. the diameter 100 ha. However, the largest active net fi eld of the pattern centre) can range from less than was only 2.3 ha. In addition, despite over 270 one metre to more than fi ve metres and the active areas were found their total cover was stony border from 0.1 m to two metres (Fig. only 28 ha. Topographical characteristics of 42). However, the average centre size is ca. active sorted net squares in relation to absent two metres. Generally the borders consist of sites are presented in Figure 44A and Table 10–40 cm stones but blocks up to 1.5 m in 16. The topographical properties differed sig- diameter may be present in thin soil sites or nifi cantly between the two groups (p < 0.01 in in large features. On level ground, slope angle Mann-Whitney U-test).

61 Results

Figure 42. Active sorted net in a temporary ponding depression (69°34’52”N, 26°13’23”E / 375 m a.s.l. / 16th of August 2002). The mesh width ranges from 40 cm to one metre.

Figure 43. Inactive sorted net from the summit of Ráššoaivi fell (69°38’15”N, 26°16’12”E / 515 m a.s.l. / 12th of August 2002). The diameter of the vegetated centre is ca. 1.5 m.

5.2.8 Sorted stripes generally from ﬁ ve to 40 cm in diameter but boulders up to one metre can be found from Sorted stripes are 0.2–2 m wide and from a a few sites. Material of the stripes is usually few metres to tens of metres long features quite rounded but some stripes can be formed (Fig. 45). The width of the non-sorted inter- of angular stones and blocks. The most ac- stripe area is often several times wider than tive stripes were without vegetation cover (Fig. the sorted parts ranging from to ten metres. 46). However, usually even the active stripes Sizes of the stones in the sorted sections are were partly covered by vegetation; stones with

62 Results

Figure 44. Mean altitude and mean slope angle of the present and absent (A) sorted net, (B) sorted stripe and (C) non-sorted soliﬂ uction terrace squares. Only active landform occurrences were used in the analyses.

Figure 45. Ca. 0.5 m wide active sorted stripes on the slope of Njávgoaivi fell (69°29’34”N, 26°19’24”E / 545 m a.s.l. / 11th of July 2004). lichen and non-sorted parts by low heath veg- várri fell has several smaller feature occurrenc- etation (see Fig. 45). es. The total area of the sorted stripes could Most of the sorted stripes occur in three not be determined because only the locations fell regions, namely Áilegas in southwest, of the features were recorded due their scat- Njávgoaivi – Suhpášoaivi area and around tered distribution. In general, the stripe areas Guivi in the northeast (Appendix 8). In Ái- are quite small, usually less than one hectare. legas, stripes are abundant on the slopes of Topographical characteristics of active sorted fells north and southwest of Davimušalaš fell. stripe squares are presented in Figure 44B and The largest fi elds in the second area are on the Table 16. The mean altitude and mean slope slopes of Njávgoaivi and Geavvogeašoaivvit angle differed signifi cantly between occur- fells. In the northeast, largest stripe fi elds can rence and non-occurrence sites (p < 0.001 in be found from the fells of Guovdoaivi, Guivi, Mann-Whitney U-test). west of Guivi and Ráššoaivi. Moreover, Ahko-

63 Results

Figure 46. Small active sorted stripe on valley slope west of Guivi fell (69°37’13”N, 26°23’11”E / 475 m a.s.l. / 25th of June 2002). The stripe is ca. 0.5 m wide and ﬁ ve metres long.

5.2.9 Non-sorted solifl uction are presented in Figure 44C and Table 16. The mean altitude and mean slope angle differed terraces signifi cantly between the present and absent sites (p < 0.001 in Mann-Whitney U-test). Non-sorted solifl uction terraces are 0.2–0.8 m high and up to a few tens of metres wide features with steep frontal riser (Figs. 26C & 47). 5.2.10 Sorted solifl uction sheets The surface gradient of the feature was often measured to be less than fi ve degrees despite Sorted solifl uction sheets are often transversal the fact that the slope above the terrace could over 50 m wide slope features (Fig. 11). They be up to 30º. The frontal riser of the feature consist of coarse material, stones, boulders can be discontinuous hummock-like (Fig. 47). and blocks. The size of the material ranges The active non-sorted solifl uction terraces from 0.2 to one metre but blocks up to two have moderate vegetation cover and steeper metres in diameter can exist. Sheets can be frontal margin than the inactive due to con- divided to two subgroups; boulder sheets and stant soil movement by solifl uction. block sheets. In the boulder sheets, the mate- Non-sorted terraces are quite common fea- rial is fairly rounded, whereas more angular tures in the study area (Appendix 9; Table 15). blocks with rather sharp edges characterise Most of the terraces are situated in northern the second type. Despite that material differ- and eastern parts of the study area. However, ences between the types occur, all sheets were smaller occurrences can be found from the grouped together. Both active and inactive fea- fell areas of Áilegas and north of Gaska-Bies- tures were identifi ed from the study area (Fig. várri. The total cover of the non-sorted ter- 48; Table 15). races could not be measured because only the The distribution of sorted solifl uction locations of the features were determined in sheets is presented in Appendix 10. The the fi eld. Topographical characteristics of ac- largest sheet areas, up to 70 ha, are found tive present non-sorted solifl uction terraces from the fell regions of Áilegas, Guovdoaivi

64 Results

Figure 47. Non-sorted soliﬂ uction terrace on the foot of 28° steep slope (69°37’20”N, 26°21’40”E / 455 m a.s.l. / 20th of June 2002). The terraces frontal riser is 35 cm high and it is partly hummocky-like.

Figure 48. Active sorted soliﬂ uction sheet with rather angular material and frontal soil embankment of ca. 40 cm high (69°28’18”N, 26°1’29”E / 400 m a.s.l. / 17th of July 2003).

– Gamoaivi, Ráššoaivi, Áhkovárri, Njávgoaivi of active present landform squares in relation – Suhpášoaivi and Suobbatoaivi. The exten- to absent sites are presented in Figure 49A and sive features are commonly boulder sheets or Table 16. The mean altitude and mean slope mixed types where smaller blocky areas are angle differed signiﬁ cantly between the two present. The largest block sheets can be found groups (p < 0.001 in Mann-Whitney U-test). from slopes of Uch-Áhkováráš fell, southwest of Guovdoaivi fell and northeast of Ruohtir fell. The total cover of the sorted soliﬂ uction sheets is 1116 ha. Topographical characteristics

65 Results

A) B) C)

Figure 49. Mean altitude and mean slope angle of the present and absent (A) sorted solifl uction sheet, (B) sorted solifl uction stream and (C) defl ation squares. Only active landform occurrences were used in the analyses.

5.2.11 Sorted soliﬂ uction streams 5.2.12 Deﬂ ations

Sorted solifl uction streams are ca. 3–20 m wide The diameter of the defl ation features range and generally from ten to several tens of me- from two metres to some tens of metres and tres long features (Fig. 50). The size of the depth from a few centimetres to ca. three me- stones range generally from 10 cm to 50 cm tres. Two main types of defl ations were iden- in diameter but boulders up to one metre can tifi ed from the study area, defl ation surfaces occur in some features. Material is often quite (Fig. 51) and defl ation depressions i.e. blow- rounded but angular stones and blocks can be outs (Fig. 17). Defl ation surfaces occur on present. In addition, some features with more glacial and glaciofl uvial soils and landform, angular material occur in the study area. The whereas depressions can only be found from active features material was only partly lichen aeolian deposits. Wind erosion surfaces were covered, whereas inactive features had fairly also observed from peaty areas, mainly from dense lichen and moss cover on the stones palsas (Fig. 4). Defl ation surfaces are usu- and blocks. ally distinctively smaller than depressions, just The sorted solifl uction streams have a quite from few centimetres to 20 cm deep and often similar distribution to the stripes and sheets less than 10 m in diameter. Only active defl a- (Appendix 11). Most of the sorted streams tion features were mapped. In this study, both occur in the fell areas of Áilegas, Guovdoaivi the identifi ed defl ation landform types, surfac- – Gamoaivi, Ráššoaivi, Áhkovárri, Njávgoaivi es and depressions, were combined together – Suohpášoaivi, Suobbatoaivi and Stuorra to ensure a suffi cient number of observations Biesvárri, where also the largest features can in the statistical analyses. be found. The cover of the sorted streams Defl ations are quite common landforms in could not be defi ned because only the loca- the study area (Appendix 12; Table 15). They tions of the features were mapped in the fi eld. are most abundant in the areas southeast and Topographical characteristics of active present east of Leakšagoađoaivi fell, on the Goike-Sit- feature squares are presented in Figure 49B nogophi esker, east and south of Ulláváráš fell, and Table 16. The mean altitude and mean on the marginal areas of Luopmošjáguotkku slope angle differed signifi cantly between the esker and generally in the southeast. Fea- occupied and unoccupied squares (p < 0.001 tures over 50 m in diameter can be found in Mann-Whitney U-test). east of Leakšagoađoaivi fell and southeast of Luopmošjáguotkku esker. However, usu-

66 Results

Figure 50. Active sorted soliﬂ uction streams on the fell slope north of Guivi fell (69°37’53”N, 26°24’48”E / 525 m a.s.l. / 23th of 2002). The width of the streams range from three to 10 m. On the right side of the photograph is part of a sorted soliﬂ uction sheet.

Figure 51. Ca. 10 m wide defl ation surface above the treeline in the northwestern part of the study area (69°38’39”N, 26°3’59”E / 400 m a.s.l. / 26th of July 2003). ally defl ations are quite small-sized and their 5.3 Distribution and abundance distribution is scattered. Therefore, only the models locations were documented. Topographical characteristics of present defl ation sites are presented in Figure 49C and Table 16. Accord- 5.3.1 Palsas ing to Mann-Whitney U-test, the occurrence squares mean slope angle differed signifi cantly The fi nal modelling data included 2272 squares (p = 0.001) from the absent sites values but of which 134 (5.9%) were occupied by palsas. mean altitude did not (p = 0.082). In univariate analyses, the largest change in deviance was caused by water cover, mean slope angle, mean of shrub cover as well as peat

67 Results and rock terrain cover (Fig. 52). The results between occupied and unoccupied squares. of the fi nal distribution model are presented However, there occurred unsuccessfully pre- in Table 17. The probability of palsa occur- dicted squares that can be seen from the Fig- rence increased with increasing peat and water ure 54. For example, two modelling squares cover and decreasing mean slope angle. The west of the Áilegas fell had very low proba- response curve of water cover was humped, bilities (0.00026 and 0.00025) but were present i.e. highest probability occurred at the inter- squares (omission error) and the square in the mediate values. The shapes of the separately Leakšagoahti valley had high probability (0.84) fi tted response curves of the variables in the but was unoccupied (commission error). Ten fi nal model are illustrated in Figure 53. The fi - commission error squares were visited to gain nal model explained ca. 52 % of the variation further information of the prediction errors in palsa distribution (Table 18). and from eight of the sites palsas were found. In model evaluation, the Kappa value of Based on the hierarchical partitioning, the 0.565 was obtained (SE = 0.064, p < 0.0001), autocovariate was the strongest predictor, which indicated a relatively good discrimina- whereas other spatial variables explained less tion capability of the model. The used opti- than one percent of the variation in the data mised cut off value was 0.32. The AUC value (Fig. 55). The most important environmental was 0.950 (SE = 0.018, p < 0.0001), which factors were peat cover, mean slope angle, indicates an excellent discrimination ability mean of topographical wetness index, shrub cover, mean altitude and proportion of concave topography. Peat and slope angle were Palsa distribution important factors in both GLM and hierarchi- 10 9 cal partitioning approaches but the effect of 8 7 water cover declined substantially in HP. 6 Of the 134 present palsa squares 111 were 5 4 used in the abundance modelling. The cover 3 2 of the palsas ranged from 0.01 ha to 9.58 ha in

Change in deviance (|x|) 1 0 the modelling squares. Mean of shrub cover, elevation-relief ratio, standard deviation of

Peat cover Water cover Mean altitude topographical wetness index, mean altitude Aspect majority Relative radiation MeanMean slope shrubangleRock cover terrain cover Mean canopyElevation-relief cover ratio Std of wetness index Mean wetnessConcave index topography and mean slope angle increased the residual Sand and gravel cover Glacigenic deposit cover deviance most in univariate analysis (Fig. 56). Figure 52. Results of the univariate analyses for pal- The variables and coefﬁ cients of the ﬁ nal sa distribution. The change in deviance is presented abundance model are presented in the Table as absolute value because the results of the calcula- 19. The cover of the palsas increased when tions were negative (see Chapter 4.2.1). peat cover increased and mean of shrub cover decrease. According to residual plots, the as-

Table 17. Variables, coefﬁ cients, standard errors (SE), z- and p-values of the ﬁ nal model for palsa distribution.

Variable Coeﬃ cient SE z p Intercept -1.550 0.566 -2.739 0.0062 Peat cover 1.744*10-3 2.463*10-4 7.081 <0.0001 Mean slope angle -1.060 0.185 -5.742 <0.0001 Water cover 1.334*10-2 3.158*10-3 4.224 <0.0001 Water cover2 -1.338*10-5 3.630*10-6 -3.686 0.0002

68 Results

Figure 53. Relationship between palsas and environmental variables. The predicted probability of palsas is derived from the logistic regression model relating each predictor separately to the response variable.

Table 18. Deviance information and degrees of freedom (D.f.) of the ﬁ nal distribution (D) and abundance (A) models.

Calibration data Null Residual Change in Change in D.f. deviance deviance deviance deviance (%) Palsa D 651.73 310.91 340.82 52.3 1585 Palsa A 29.315 18.440 10.875 37.1 75 Convex non-sorted circle D 1033.76 683.97 349.79 33.8 1584 Convex non-sorted circle A 93.254 56.908 36.346 39.0 151 Stony earth circle D 1094 1010.4 83.6 7.6 1584 Earth hummock D 2001.2 1060.4 940.8 47.0 1582 Earth hummock A 671.92 323.24 348.68 51.9 1026 Peat pounu D 1928.8 1024.3 904.5 46.9 1578 Peat pounu A 303.91 160.33 143.58 47.2 426 Stone pit D 1055.6 856.9 198.7 18.8 1584 Sorted net D 933.55 801.42 132.13 14.2 1584 Sorted stripe D 904.93 509.21 395.72 43.7 1586 Non-sorted solifl uction terrace D 1131.3 1017.2 114.1 10.1 1586 Sorted solifl uction sheet D 1467.65 584.96 882.69 60.1 1582 Sorted solifl uction sheet A 226.13 123.03 103.1 45.6 278 Sorted solifl uction stream D 1024.95 561.24 463.71 45.2 1587 Defl ation D 1210.72 962.36 248.36 20.5 1583 sumption of normal errors was appropriate x-coordinate 1.1% of the variance in the data despite that a few possible outliers were present (Fig. 58). Peat cover, mean of topographical (Fig. 57). The fi nal abundance model explained wetness index, mean slope angle, water cover 37% of the variation in the data (Table 18). and proportion of concave topography were in-

Spearman’s rank correlation coefﬁ cient (Rs) dependently the most inﬂ uential environmental for the evaluation data was 0.2739 (p < 0.0001) correlates. Peat cover was the most important which was substantially less than for the cali- factor in both statistical approaches but the ef- bration data (Rs = 0.6470, p < 0.0001). Autoco- fect of the shrub cover decreased drastically in variate for palsa abundance explained 5.4% and the hierarchical partitioning analysis.

69 Results

Figure 54. Observed and predicted distribution of palsas in the whole modelling area. The probabilities of palsa occurrence were calculated for the whole modelling area with a calibrated logistic regression model at 25-ha resolution. This palsa model is an example where the Kappa measure (0.565) indicated a good, and the area under the curve (AUC) value (0.950) an excellent, discrimination ability of the model.

Palsa distribution

6 5 4 3 2 1 0 Figure 55. Results of the hierarchical partitioning Explained variance (%) analyses for palsa distribution. The partitioning approach included nine environmental and three spa- Peat cover x-coordinatey-coordinateWater cover Autocovariate Mean altitude tial variables. The independent contribution of the Mean slopeMean angle shrub cover Mean canopy cover predictor is given as a percentage of the total vari- Mean wetness indexConcave topographyStd of wetness index ance (Chapter 4.2.3).

70 Results

Table 19. Variables, coefﬁ cients, standard errors (SE), t- and p-values of the ﬁ nal model for palsa abundance.

Variable Coeﬃ cient SE t p Intercept -0.817 0.138 -5.930 <0.0001 Peat cover 8.365*10-4 1.288*10-4 6.494 <0.0001 Mean shrub cover -3.253*10-2 9.119*10-3 -3.568 0.0006

Palsa abundance Palsa abundance

1.4 6 1.2 5 1 0.8 4 0.6 3 0.4 2 0.2 1

Change in deviance (|x|) 0 0 Explained variance (%)

Peat cover Water cover Mean altitude Peat cover Aspect majority Water coverx-coordinate y-coordinate Mean shrub cover Mean slope angle Relative radiation Autocovariate Mean altitude Elevation-relief ratio Mean canopy cover Rock terrain cover Std of wetness index ConcaveMean topography wetness index Sand and gravel cover Mean slope angle Mean shrub cover Glacigenic deposit cover Mean canopy cover Mean wetness indexConcave topography Std of wetness index

Figure 56. Results of the univariate analyses for Figure 58. Results of the hierarchical partitioning palsa abundance (for more details see Fig. 52). analyses for palsa abundance (for more details see Fig. 55).

Figure 57. Evaluation of the fi nal calibration model for the palsa abundance with regression residuals: (A) residuals vs fi tted values and (B) normal Q–Q plot. Residuals should not show systematic patterns in fi gure (A) and residuals should be approximately normally distributed i.e. they should form rather straight line in fi gure (B).

tion, mean of canopy cover as well as mean 5.3.2 Convex non-sorted circles of topographical wetness index were the most The number of present convex non-sorted important variables for the circle distribution circle squares used in the modelling was 221 (Fig. 59). It is notable that the effect of mean (9.7%), whereas 2051 squares were unoccu- altitude was over threefold compared to the pied. According to univariate analyses mean second most important variable, mean slope altitude, mean slope angle, relative solar radia- angle. The results of the binary GLM are pre-

71 Results

Table 20. Variables, coefﬁ cients, standard errors (SE), z- and p-values of the ﬁ nal model for convex non-sorted circle distribution.

Variable Coeﬃ cient SE z p Intercept 43.11 7.296 -5.908 <0.0001 Mean slope angle -0.692 7.908*10-2 -8.748 <0.0001 Concave topography 7.095*10-2 9.985*10-3 7.105 <0.0001 Mean altitude 0.209 3.313*10-2 6.297 <0.0001 Mean altitude2 -2.115*10-4 3.896*10-5 -5.431 <0.0001 Relative radiation -0.113 3.009*10-2 -3.745 0.0002

Convex non-sorted circle distribution presented in Figure 61. Most of the present

60 squares with low predicted probability (< 0.2) 50 were outside the core distribution area, name- 40

30 ly in the western and southern parts of the 20 Báišduattar – Áilegas area, and almost all ab- 10 sent squares with high predicted probability (>

Change in deviance (|x|) 0 0.6) were in the northeast and east. Seventeen

Peat cover of the commission error squares were visited Mean altitude Water cover Aspect majority Relative radiation Mean slope angle Mean shrub coverRock terrain cover and six of them were found to contain convex Mean canopy cover Elevation-relief ratio MeanConcave wetness index topographyStd of wetness index Sand and gravel cover Glacigenic deposit cover non-sorted circles. The autocovariate variable explained 8.3% Figure 59. Results of the univariate analyses for of the variance in the data, which is over two- convex non-sorted circle distribution (for more de- fold more than the most important environ- tails see Fig. 52). mental variable, namely mean of canopy cover (Fig. 62). Furthermore, the y-coordinate ex- sented in Table 20. The probability of convex plained 3.1% and x-coordinate 1.7% of the non-sorted circles occurrence increased when variance in the data. In addition to mean of the proportion of concave topography as well canopy cover, the most infl uential environ- as mean altitude increased and mean slope mental variables were mean altitude, mean of angle and relative radiation decreased. The topographical wetness index, mean slope angle response curve of mean altitude expressed a and proportion of concave topography. Mean non-linear association between the distribu- of canopy cover and mean of topographical tion of convex non-sorted circles and altitude, wetness index were not in the fi nal distribution and the shape of the curve is demonstrated in model. Furthermore, the independent contri- Figure 60. The fi nal model explained 33.8% of bution of relative radiation was very low in the deviance change in the data (Table 18). HP despite the variable was highly signifi cant The evaluation measures calculated with in the distribution model. the evaluation data indicated good discrimina- Abundance modelling of convex non- tion ability of the logistic regression model. sorted circles was conducted with 221 present The Kappa value was 0.482 (SE = 0.056, p < squares. The cover of the feature ranged from 0.001) and AUC of ROC plot was 0.879 (SE 0.01 ha to 13.49 ha in the modelling squares. = 0.023, p < 0.0001). The Kappa measure Mean altitude, mean of topographical wetness was calculated with cut off value of 0.31. The index, mean slope angle, aspect majority and calibration model was utilised to predict the standard deviation of wetness index had the occurrence of convex non-sorted circle for greatest effect on the deviance change in uni- the whole modelling area and the results are variate analyses (Fig. 63). The difference be-

72 Results

Convex non-sorted circle distribution

9 8 7 6 5 4 3 2 1 0 Explained variance (%)

Peat cover y-coordinate x-coordinate Autocovariate Mean altitude Mean slope angle Mean shrub cover Relative radiation Mean canopy cover Mean wetness indexConcave topography Glacigenic deposit cover

Figure 60. Relationship between the convex non- Figure 62. Results of the hierarchical partitioning sorted circle occurrence and mean altitude (for analyses for convex non-sorted circle distribution more details see Fig. 53). (for more details see Fig. 55).

Figure 61. Observed and predicted distribution of convex non-sorted cirlces in the whole modelling area (for more details see Fig. 54). The model is an example where the Kappa measure (0.482) indicated rather good, and the area under the curve (AUC) value (0.879) good, discrimination ability of the ﬁ nal model.

73 Results

(Fig. 65). In addition, both coordinate variables Convex non-sorted circle abundance explained over three percent of the variance in 16 14 the response data. Four environmental vari- 12 10 ables, namely mean of topographical wetness 8 6 index, mean altitude, proportion of concave 4 topography and mean slope angle, explained 2

Change in deviance (|x|) 0 over two percent and all other environmental predictors explained over one percent of the

Peat cover Mean altitude Water cover variance in the data. Therefore, all variables in Aspect majority Mean slope angle Mean shrub coverRelative radiation Rock terrain cover Elevation-relief ratio Mean canopy cover Mean wetness indexStd of wetness index Concave topography the fi nal GLM model were also important ac- Sand and gravel cover Glacigenic deposit cover cording to HP. Figure 63. Results of the univariate analyses for convex non-sorted circle abundance (for more details see Fig. 52). 5.3.3 Stony earth circles tween mean altitude and the second most im- A total of 243 (10.7%) stony earth circle portant variable was even more striking than squares were used in the modelling. Mean al- in the distribution data; the effect of mean titude, mean of topographical wetness index, altitude was over four-fold compared to the mean of canopy cover, relative solar radiation, moisture index. proportion of concave topography and mean The results of the GLM showed that the of shrub cover caused the largest changed in abundance of convex non-sorted circles in- deviance in univariate analyses (Fig. 66). The creased when mean altitude as well as mean variables and coeffi cients of the fi nal logistic of topographical wetness index increased and regression model for feature distribution are mean slope angle decreased (Table 21). The presented in Table 22. The probability of fea- assumption of normal errors was adequate ture occurrence increased with increasing gla- because residuals were distributed relatively cigenic deposit cover as well as mean altitude evenly and, above all, they were normally dis- and decreasing mean of topographical wetness tributed (Fig. 64). The only discernible pattern index and mean of canopy cover. The humped in the residuals is caused by the landform clas- response curve of mean altitude is displayed in sifi cation system. Pattern areas, which were Figure 67. It is notable that the fi nal model ex- mapped by single points, can be detected from plained only 7.6% of the variation in the stony the residual plot (Fig. 64). The fi nal abundance earth circle data (Table 18). model explained 39% of the deviance change Cohen’s Kappa statistic for evaluation data in the data (Table 18). indicated poor discrimination power of the In the model evaluation, the Spearman’s model (κ = 0.254, SE = 0.053, p < 0.0001, cut rank correlation coeffi cient was 0.650 (p < off value = 0.20). The area under the curve 0.0001) that was even better than for the cali- measure was 0.735 (SE = 0.033, p < 0.0001) bration data (Rs = 0.633, p < 0.0001). Autoco- that indicated a rather good performance of variate had the greatest effect on the response the model. However, the evaluation value was

Table 21. Variables, coefﬁ cients, standard errors (SE), t- and p-values of the ﬁ nal model for convex non-sorted circle abundance.

Variable Coeﬃ cient SE t p Intercept -8.619 0.991 -8.694 <0.0001 Mean altitude 1.072*10-2 1.318*10-3 8.132 <0.0001 Mean wetness index 0.397 7.553*10-2 5.258 <0.0001 Mean slope angle -0.126 3.721*10-2 -3.387 <0.0001

74 Results

Figure 64. Evaluation of the ﬁ nal calibration model for the convex non-sorted circle abundance with regression residuals: (A) residuals vs ﬁ tted values and (B) normal Q–Q plot (for more details see Fig. 57).

Convex non-sorted circle abundance Stony earth circle distribution

30 8 7 25 6 20 5 4 15 3 10 2 1 5 0 0 Change in deviance (|x|) Explained variance (%)

Peat cover Mean altitude Water cover y-coordinatex-co ordinate Aspect majority AutocovariateMean altitude Relative radiation Mean shrubMean cover slope angle Rock terrain cover Aspect majority Mean canopy cover Elevation-relief ratio Mean wetness indexConcave topography Std of wetness index Mean slope angle Mean shrub cover Sand and gravel cover Mean wetness index Concave topography Glacigenic deposit cover Std ofSand wetness and indexgravel cover Glacigenic deposit cover Figure 66. Results of the univariate analyses for Figure 65. Results of the hierarchical partitioning stony earth circle distribution (for more details see analyses for convex non-sorted circle abundance Fig. 52). (for more details see Fig. 55). fairly close to the limit of poor model. The of shrub cover, topographical wetness index relatively poor discrimination ability of the and canopy cover were the next most impor- model can also be detected from the predic- tant environmental correlates. The results of tion map (Fig. 68). However, the prediction the two statistical approaches differed partly; accuracy of the model was slightly better than mean of shrub cover was not in the GLM indicated by Kappa measure in the core dis- model and the independent effect of mean tribution areas because all of the ﬁ ve visited altitude on the feature distribution was low in absent squares with relatively high (> 0.2) pre- HP analysis. dicted probability of occurrence contained stony earth circles. In hierarchical partitioning, the autocovari- 5.3.4 Earth hummocks ate explained ca. 16.9% of the variance in the data, which was over four-fold more than the The distribution modelling of active earth most important environmental variable, name- hummocks was performed with 1553 (68.4%) ly cover of glacigenic deposit (Fig. 69). Mean present squares. Mean of shrub cover, peat

75 Results

Table 22. Variables, coefﬁ cients, standard errors (SE), z- and p-values of the ﬁ nal model for stony earth circle distribution.

Variable Coeﬃ cient SE z p Intercept -11.59 5.056 -2.293 0.0219 Glacigenic deposit cover 7.330*10-4 1.787*10-4 4.102 <0.0001 Mean wetness index -0.474 0.125 -3.799 0.0001 Mean canopy cover -5.228*10-2 1.388*10-2 -3.768 0.0002 Mean altitude 7.568*10-2 2.496*10-2 3.032 0.0024 Mean altitude2 -1.077*10-4 3.156*10-5 -3.413 0.0006

explained 47% of the variation in the response data (Table 18). The discrimination power of the model was good (κ = 0.637, SE = 0.032, p < 0.0001) or excellent (AUC = 0.920, SE = 0.011, p < 0.0001) depending on the utilised evaluation measure. The optimal Kappa statistic was obtained with cut of value of 0.56. According to the fi eld checking, the distribution model was even better than indicated by the evaluation values because fi ve of the ten absent squares with high predicted probability (> 0.8) were present squares. In general, unsuccessful predictions were distributed relatively randomly in the modelling area (Fig. 72). However, most Figure 67. Relationship between stony earth circles of the commission errors were in or near the and mean altitude (for more details see Fig. 53). core distribution area, whereas omission errors (predicted probability < 0.2) were in the distributional margins i.e. in fell regions. cover, mean altitude, standard deviation as well Based on hierarchical partitioning, auto- as mean of topographical wetness index and covariate was the best explanatory variable water cover were the most infl uential environ- (Fig. 73). Seven of the nine environmental mental variables in the univariate analyses (Fig. predictors explained over one percent of the 70). Shrub and peat cover were distinguished variance in the response data, namely mean of from the set of predictors because their effect shrub cover, peat cover, mean altitude, mean was over two-fold compared to the other vari- of topographical wetness index, proportion ables. The fi nal distribution model included of concave topography, standard deviation of fi ve explanatory variables (Table 23). The topographical wetness index and mean slope probability of fi nding earth hummocks from angle. Thus, all fi ve variables of the fi nal distri- the modelling square increased when peat bution model were among the six most infl u- cover, mean of shrub cover, mean as well as ential factors in the partitioning routine. standard deviation of topographical wetness Abundance modelling was performed with index increased and mean altitude decreased. 1478 active present squares. The cover of the The association between response and peat as earth hummocks ranged from 0.01 ha to 25 well as shrub cover was non-linear. The shape ha in the data, i.e. some of the squares were of the response curve of peat cover, shrub fully occupied by earth hummocks. Univariate cover and topographical wetness index is il- analyses indicated that mean of shrub cover is lustrated in Figure 71. The fi nal logistic model a superior correlate for earth hummock abun-

76 Results

Figure 68. Observed and predicted distribution of stony earth circle in the whole modelling area (for more details see Fig. 54). Stony earth circle model is an example where the Kappa measure (0.254) indicated poor, and the area under the curve (AUC) value (0.735) fairly good, discrimination ability of the model.

Stony earth circle distribution

18 16 14 12 10 8 6 4 2 0 Explained variance (%)

y-coordinate x-coordinate Autocovariate Mean altitude Aspect majority Mean shrub coverMean slope angle Relative radiation Mean canopy cover Figure 69. Results of the hierarchical partitioning Mean wetness indexConcave topography Glacigenic deposit cover analyses for stony earth circle distribution (for more details see Fig. 55).

77 Results

Earth hummock distribution mocks increased when mean of shrub cover, 60

50 peat cover and mean of topographical wetness 40 index increased and mean slope angle, mean 30 altitude and water cover decreased (Table 24). 20

10 The effect of shrub and peat cover was non-

Change in deviance (|x|) 0 linear as in the distribution model. The assumption of normal errors was fairly proper

Peat cover Mean altitude Water cover but, again, the grid approach and the features Aspect majority Relative radiation Mean shrub cover Rock terrain cover Mean slope angle Mean canopy cover Elevation-relief ratio Std ofMean wetness wetness indexConcave index topography Sand and gravel cover size classifi cation system caused some minor Glacigenic deposit cover trend in the residual plot. Six predictors of the Figure 70. Results of the univariate analyses for fi nal model explained ca. 52% of the deviance earth hummock distribution (for more details see change in the data (Table 18). Fig. 52). Spearman’s rank correlation coeffi cient for the evaluation data set was 0.722 (p < 0.0001) dance (Fig. 74). Other potentially important that indicated a relatively high prediction abil- but clearly less effective factors are peat cover, ity. According to the hierarchical partition- mean slope angle, mean altitude, water cover ing, autocovariate had the highest explana- and the mean as well as standard deviation of tion power but the distinction with the most topographical wetness index. important environmental factor, shrub cover, The results of the fi nal abundance model was in fact negligible (Fig. 75). In addition, showed that the abundance of earth hum- peat cover, mean of topographical wetness

Table 23. Variables, coeffi cients, standard errors (SE), z- and p-values of the fi nal model for earth hummock distribution. Variable Coeffi cient SE z p Intercept -7.209 1.452 -4.966 <0.0001 Peat cover 1.211*10-2 1.672*10-3 7.241 <0.0001 Peat cover2 -5.550*10-6 7.216*10-7 -7.692 <0.0001 Mean wetness index 0.865 0.122 7.121 <0.0001 Mean shrub cover 0.258 3.697*10-2 6.988 <0.0001 Mean shrub cover2 -6.727*10-3 3.966*10-3 -3.966 <0.0001 Mean altitude -1.097*102 1.578*10-3 -6.954 <0.0001 Std of wetness index 1.066 0.202 5.273 <0.0001

Figure 71. Relationship between earth hummocks and environmental variables (for more details see Fig. 53).

78 Results

Figure 72. Observed and predicted distribution of earth hummock in the whole modelling area (for more details see Fig. 54).

Earth hummock distribution Earth hummock abundance

60 4 3.5 50 3 40 2.5 2 30 1.5 20 1 10 0.5

0 Change in deviance (|x|) 0 Explained variance (%)

Peat cover Peat cover Mean altitudeWater cover y-coordinatex-coordinate Aspect majority Autocovariate Mean altitude Relative radiation Mean shrubMean cover slope angle Rock terrain cover Elevation-relief ratio Mean canopy cover Mean shrub cover Mean slope angle MeanStd wetness of wetness indexConcave index topography Mean canopy cover Sand and gravel cover Mean wetnessConcaveStd index topographyof wetness index Glacigenic deposit cover Glacigenic deposit cover Figure 74. Results of the univariate analyses for Figure 73. Results of the hierarchical partitioning earth hummock abundance (for more details see analyses for earth hummock distribution (for more Fig. 52). details see Fig. 55).

79 Results

Table 24. Variables, coefﬁ cients, standard errors (SE), t- and p-values of the ﬁ nal model for earth hummock abundance.

Variable Coeffi cient SE t p Intercept -1.207 0.330 -3.659 0.0003 Mean shrub cover 6.765*10-2 6.464*10-3 10.467 <0.0001 Mean shrub cover2 -1.281*10-3 2.138*10-4 -5.993 <0.0001 Peat cover 1.031*10-3 1.053*10-4 9.790 <0.0001 Peat cover2 -5.182*10-7 5.541*10-8 -9.351 <0.0001 Mean slope angle -5.988*10-2 9.240*10-3 -6.480 <0.0001 Mean wetness index 0.158 2.982*10-2 5.298 <0.0001 Mean altitude -1.506*10-3 4.234*10-4 -3.556 0.0004 Water cover -3.590*10-3 1.026*10-3 -3.500 0.0005 index, mean slope angle, glacigenic deposit, altitude, relative solar radiation and mean of proportion of concave topography and mean topographical wetness index stand out from altitude were the most infl uential factors in the the rest of the variables. The set of the predic- partitioning routine. The water cover variable tors in the fi nal logistic regression model was was signifi cant in the fi nal GLM model but its diverse, including eight correlates. The prob- independent effect on the response was fairly ability of peat pounu occurrence increased marginal in HP, ca. 0.5%. when the cover of peat, water and shrub as well as standard deviation of topographical wetness index, mean of canopy cover and pro- 5.3.5 Peat pounus portion of concave topography increased and mean slope angle and proportion of relative Active peat pounus were located from 678 radiation decreased (Table 25). Three of the (29.8%) modelling squares. According to uni- predictors were non-linearly associated with variate analyses, mean slope angle, water cover, the response, namely peat and water cover as peat cover, mean of shrub cover, proportion well as relative solar radiation. The shape of of concave topography, standard deviation the response curve of peat cover, mean slope of topographical wetness index and mean of angle and shrub cover is shown in Figure 77. canopy cover have the greatest effect on the The fi nal model explained ca. 47% of the devi- feature distribution (Fig. 76). Moreover, mean ance change in the pounu data (Table 18).

Earth hummock abundance

6 Peat pounu distribution 5 35 4 30 3 25 2 20 1 15 0 10

Explained variance (%) 5 0 Change in deviance (|x|)

Peat cover y-coordinatex-coordinateWater cover Autoco variate Mean altitude

Mean shrub cover Mean slope angle Peat cover Water cover Mean wetness indexConcave topography Mean altitude Std of wetness index Aspect majority Relative radiation Mean slope angleMean shrub cover Rock terrain cover Glacigenic deposit cover Mean canopy cover Elevation-relief ratio ConcaveStd topography of wetness index Mean wetness index Sand and gravel cover Glacigenic deposit cover Figure 75. Results of the hierarchical partitioning analyses for earth hummock abundance (for more Figure 76. Results of the univariate analyses for peat details see Fig. 55). pounu distribution (for more details see Fig. 52).

80 Results

The Kappa value was 0.641 (SE = 0.032, In HP analysis, the autocovariate and peat p < 0.001) and AUC of ROC plot was 0.913 cover explained about the same amount of (SE = 0.012, p < 0.0001). Therefore, the dis- variance of the response data (Fig. 79). The crimination ability of the distribution model other most inﬂ uential environmental factors was good or excellent depending on the uti- were mean of shrub cover, mean slope angle, lised evaluation approach. The κ measure was proportion of concave topography and mean calculated with the cut off value of 0.41. The as well as standard deviation of topographi- probability map of peat pounu occurrence for cal wetness index. Explanation power of wa- the whole modelling area is presented in Fig- ter cover, relative solar radiation and mean ure 78. Most of the present squares with low of canopy cover was distinctively less than pounu probability (< 0.1) were in the eastern, one percent despite that they were in the ﬁ nal and absent squares with high probability (> GLM model. 0.8) were in the southern, part of the study Abundance modelling was conducted with area and around the Áitečohkka fell. Twelve 615 present pounu squares. The cover of the of the commission error squares were visited feature ranged from 0.01 ha to 22.9 ha in the and all of them were found to contain peat modelling squares. Mean slope angle had the pounus. greatest effect on pounu abundance in univari-

Table 25. Variables, coefﬁ cients, standard errors (SE), z- and p-values of the ﬁ nal model for peat pounu distribution.

Variable Coeﬃ cient SE z p Intercept 88.35 25.41 3.477 0.0005 Peat cover 4.698*10-3 4.705*10-4 9.985 <0.0001 Peat cover2 -1.635*10-6 2.488*10-7 -6.572 <0.0001 Mean slope angle -0.430 6.450*10-2 -6.665 <0.0001 Std of wetness index 0.999 0.221 4.524 <0.0001 Mean canopy cover 3.845*10-2 9.321*10-3 4.125 <0.0001 Water cover 5.723*10-3 1.398*10-3 4.093 <0.0001 Water cover2 -2.818*10-6 7.679*10-7 -3.669 0.0002 Concave topography 3.685*10-2 9.497*10-3 3.880 0.0001 Relative radiation -2.315 0.6279 -3.687 0.0002 Relative radiation2 1.423*10-2 3.874*10-3 3.674 0.0002 Mean shrub cover 4.253*10-2 1.175*10-2 3.620 0.0003

Figure 77. Relationship between peat pounu distribution and environmental variables (for more details see Fig. 53).

81 Results

Figure 78. Observed and predicted distribution of peat pounus in the whole modelling area (for more details see Fig. 54).

Peat pounu distribution Peat pounu abundance

8 6 7 5 6 5 4 4 3 3 2 2 1 1

Change in deviance (|x|) 0 0 Explained variance (%)

Peat cover Water cover Mean altitude Aspect majority Mean slope angle Relative radiationMean shrub cover Peat cover Rock terrainMean cover canopy cover Elevation-relief ratio Waterx-coordinate covery-co ordinate Mean wetnessConcave index topographyStd of wetness index Autocovariate Sand and gravel cover Glacigenic deposit cover Mean shrubMean coverslope angle Relative radiation Mean canopy cover ConcaveMean topographyStd wetn of ess wetness index index Figure 80. Results of the univariate analyses for peat Figure 79. Results of the hierarchical partitioning pounu abundance (for more details see Fig. 52). analyses for peat pounu distribution (for more details see Fig. 55).

82 Results ate analyses and the deviance change ability of the model’s ability to predict pounu abundance the other variables was over two-fold less (Fig. with evaluation data was good (Rs = 0.715, p

80). It is notable that peat was in tenth position < 0.0001), (3) the model was rather robust (Rs and such variables as sand and gravel, rock of calibration model was 0.689) and (4) the terrain and relative solar radiation were more results of the GLM model concorded fairly infl uential. The results of the fi nal model are well with the outcomes of HP analyses (Fig. presented in Table 26. The pounu abundance 82). The peat cover explained 6.7%, mean increased when peat cover increased and mean slope angle 3.4%, whereas elevation-relief ra- slope angle as well as elevation-relief ratio de- tio explained only 0.5% of the variation in the creased in the modelling square. The non-line- data. Other notable correlates were mean of ar association between pounus and slope angle topographical wetness index, mean of shrub is demonstrated in Figure 81. The fi nal abun- cover and proportion of concave topography. dance model explained 47.2% of the variation Finally, it is worth of noting that, for the fi rst in the data (Table 18). time in this study, the environmental variable The residuals of the calibrated model had explained more than the autocovariate; how- a slight trend and their distribution was only ever, the difference was not great. fairly close to normal. However, the calibration model was accepted because: (1) the utilised transformation gave the best result, (2)

Table 26. Variables, coefﬁ cients, standard errors (SE), t- and p-values of the ﬁ nal model for peat pounu abundance.

Variable Coeﬃ cient SE t p Intercept 0.544 0.204 2.662 0.0081 Peat cover 6.695*10-4 5.750*10-5 11.643 <0.0001 Mean slope angle -0.377 6.297*10-2 -5.987 <0.0001 Mean slope angle2 1.444*10-2 6.836*10-3 4.600 <0.0001 Elevation-relief ratio -1.007 0.284 -3.549 0.0004

Peat pounu abundance

7 6 5 4 3 2 1 0 Explained variance (%)

Peat cover x-coordinate y-coordinate Autocovariate Mean altitude Mean slope angleMean shrub cover Elevation-relief ratio Mean wetnessConcave index topography Sand and gravel coverStd of wetness index

Figure 81. Relationship between peat pounu abun- Figure 82. Results of the hierarchical partitioning dance and mean slope angle (for more details see analyses for peat pounu abundance (for more de- Fig. 53). tails see Fig. 55).

83 Results

5.3.6 Stone pits of the deviance change in the data (Table 18). Kappa value of the evaluation data was Of the 2272 modelling squares 247 (10.9%) 0.318 (SE = 0.047, p < 0.0001, cut off value were found to contain active stone pits. Ac- = 0.19) indicating fairly poor discrimination cording to univariate analyses, standard de- ability of the model. Instead, area under the viation of topographical wetness index, mean curve (AUC) measure of 0.806 (SE = 0.024, slope angle, mean of shrub cover, mean of p < 0.0001) indicated rather good model per- topographical wetness index, as well as sand formance. Predicted probabilities of stone pit and gravel cover have the greatest effect on occurrence for the whole modelling area are stone pit occurrence (Fig. 83). Based on lo- visualised in Figure 85. Present squares have, gistic regression, the distribution of features in general, higher predicted probabilities than is determined by mean slope angle, mean of absent squares, which concords well with the shrub cover, standard deviation of topograph- threshold independent evaluation measure. ical wetness index and moraine cover (Table When the probabilities are truncated back to 27). The shape of the response curve of glasi- binary form, the worse modelling perform- genic deposit and slope angle are exhibited in ance is revealed. However, based on ﬁ eld Figure 84. The ﬁ nal model explained ca. 19% evaluation the distribution model is still better than indicated by Kappa measure because 13

Stone pit distribution of the 19 absent squares with relatively high predicted probability (> 0.3) were occupied by 12 10 stone pits. 8 In hierarchical partitioning, the autocovari- 6

4 ate variable explained independently ca. 4.1% 2 of the variation in the response data (Fig. Change in deviance (|x|) 0 86). Eight of the environmental predictors explained more than one percent of the vari- Peat cover Mean altitude Water cover Aspect majority ance, namely mean slope angle, mean of shrub MeanMean slope shrub angle coverRelative radiation Mean canopy cover Rock terrain coverElevation-relief ratio Std of wetness indexMean wetness index Concave topography Sand and gravel cover Glacigenic deposit cover cover, mean of topographical wetness index, peat cover, standard deviation of topographi- Figure 83. Results of the univariate analyses for cal wetness index, proportion of concave to- stone pit distribution (for more details see Fig. 52). pography, mean altitude and moraine cover.

Figure 84. Relationship between stone pits and glacigenic deposit cover and mean slope angle (for more details see Fig. 53).

84 Results

Figure 85. Observed and predicted distribution of stone pits in the whole modelling area (for more details see Fig. 54).

Table 27. Variables, coefﬁ cients, standard errors (SE), z- and p-values of the ﬁ nal model for stone pit distribution.

Variable Coeﬃ cient SE z p Intercept -5.209 0.939 -5.548 <0.0001 Mean slope angle -0.393 6.669*10-2 -5.885 <0.0001 Mean shrub cover 5.796*10-2 1.067*10-2 5.431 <0.0001 Std of wetness index 1.004 0.264 3.796 0.0001 Glacigenic deposit cover 2.011*10-3 5.702*10-4 3.527 0.0004 Glacigenic deposit cover2 -6.478*10-7 1.920*10-7 -3.375 0.0007

85 Results

Stone pit distribution Sorted net distribution

25 4.5 4 20 3.5 3 15 2.5 10 2 1.5 5 1

0.5 Change in deviance (|x|) 0 0 Explained variance (%)

Peat cover Mean altitude Water cover Aspect majority Peat cover MeanMean slope shrub angle cover Relative radiation y-coordinatex-coordinate Rock terrain coverMean canopy coverElevation-relief ratio Autocovariate Mean altitude Std of wetness indexMean wetness index Concave topography Sand and gravel cover MeanMean slope angleshrub cover Glacigenic deposit cover Mean wetness indexConcave topography Std of wetness index Sand and gravel cover Glacigenic deposit cover Figure 87. Results of the univariate analyses for sorted net distribution (for more details see Fig. Figure 86. Results of the hierarchical partitioning 52). analyses for stone pits distribution (for more details see Fig. 55). The humped shape of the response curve of peat cover is illustrated in Figure 88. The fi - Slope and shrub variables were important nal model explained ca. 14% of the deviance predictors based on both approaches but the change in the data (Table 18). standard deviation of topographical wetness In model evaluation, a Kappa measure of index and glacigenic deposit were clearly less 0.170 (SE = 0.053, p < 0.0001, cut off value effective in HP than in GLM. = 0.18) and an AUC value of 0.722 (SE = 0.035, p < 0.0001) were obtained. The κ measure indicated poor model performance and 5.3.7 Sorted nets concorded well with the low proportion of explained deviance of the calibration model. The number of present active sorted net The fairly poor prediction ability of the model squares used in the modelling was 185 (8.1%). can also be seen in Figure 89. Once again, the Standard deviation of topographical wetness threshold-independent measure indicated a index, mean altitude, peat cover and mean of better discrimination capability of the distri- topographical wetness index changed most the bution model than the threshold-dependent deviance in univariate analyses (Fig. 87). The approach. However, the AUC value was rel- same set of correlates was in the fi nal logistic atively close to the limit of poor model; the regression model (Table 28). The probabil- difference was less then the amount of one ity of fi nding sorted net from the modelling standard error. Despite that the evaluation re- square increased when the standard deviation sults indicated insuffi cient prediction ability of and mean of topographical wetness index, the model, it can be used to detect new feature mean altitude as well as peat cover increased. areas because six of the seventeen commis-

Table 28. Variables, coeffi cients, standard errors (SE), z- and p-values of the fi nal model for sorted net distribution. Variable Coeffi cient SE z p Intercept -21.92 2.062 -10.632 <0.0001 Std of wetness index 1.753 0.252 6.945 <0.0001 Mean wetness index 0.968 0.143 6.790 <0.0001 Mean altitude 1.276*10-2 2.128*10-3 5.997 <0.0001 Peat cover 1.860*10-3 6.058*10-4 3.070 0.0021 Peat cover2 -1.318*10-6 3.725*10-7 -3.537 0.0004

86 Results

sion error sites (predicted probability > 0.3) were found to contain small ﬁ elds of sorted nets. Based on the HP, the autocovariate had the highest contribution in the sorted net occurrence (Fig. 90). Independently the most inﬂ u- ential environmental factors were proportion of concave topography, mean of topographical wetness index, mean slope angle and standard deviation of topographical wetness index. In addition, water cover, mean altitude, peat cover and glacigenic deposit cover explained over one percent of the variation in the sorted Figure 88. Relationship between sorted nets and net data. The differences between GLM and peat cover (for more details see Fig. 53). HP results were quite clear because only the

Figure 89. Observed and predicted distribution of sorted nets in the whole modelling area (for more details see Fig. 54).

87 Results

Sorted net distribution Sorted stripe distribution

60 10 9 50 8 40 7 6 30 5 4 20 3 2 10 1 0 Change in deviance (|x|) 0 Explained variance (%)

Peat cover Mean altitude Water cover Peat cover Aspect majority y-coordinateWater cover x-coordinate Mean slope angle Mean shrub cover Relative radiation Autocovariate Mean altitude Mean canopy cover Rock terrain cover Elevation-relief ratio Concave topography MeanStd wetness of wetness index index Sand and gravel cover Mean slope angle Mean shrub cover Glacigenic deposit cover ConcaveMean topography wetnessStd index of wetness index Glacigenic deposit cover Figure 91. Results of the univariate analyses for Figure 90. Results of the hierarchical partitioning sorted stripe distribution (for more details see Fig. analyses for sorted net distribution (for more details 52). see Fig. 55). wetness index variables were among the most based on the area under the curve measure important factors in partitioning. (AUC±SE = 0.937±0.012, p < 0.0001). The good success of the model can also be seen in the prediction map (Fig. 93). Nevertheless, 5.3.8 Sorted stripes over thirty of the present squares had low predicted probability (less than 0.1) and twenty The total number of the present active stone of the squares with fairly high probability (> stripe squares used in the modelling was 185 0.6) were absent squares. However, in the fi eld (8.1%). Mean slope angle and mean altitude checking, the model was shown to be slightly were distinctively the most infl uential predic- better because three of the commission error tors in univariate analyses (Fig. 91) and these squares (predicted probability > 0.8) were ac- factors were the only variables in the fi nal dis- tually present sites. tribution model as well (Table 29). The prob- Mean altitude explained 6.1% of the variability of feature presence increased when ation in the sorted stripe data and was inde- mean altitude and mean slope angle increased. pendently measured as the most important The shape of the response curve of altitude variable in hierarchical partitioning (Fig. 94). was s-shaped and slope angle humped (Fig. However, the explanation power of the au- 92). The fi nal distribution model explained ca. tocovariate was fairly close to the proportion 44% of the deviance change in the response of mean altitude. Moreover, the y-coordinate data (Table 18). explained slightly over one percent of the vari- According to the evaluation, the predic- ance. Mean slope angle, mean of shrub cover, tion ability of the model was good based on standard deviation of topographical wetness the Kappa value (κ±SE = 0.437±0.068, p < index, glacigenic deposit cover and proportion 0.0001, cut off value = 0.32) and excellent of concave topography had clearly less expla-

Table 29. Variables, coefﬁ cients, standard errors (SE), z- and p-values of the ﬁ nal model for sorted stripe distribution.

Variable Coeﬃ cient SE z p Intercept -18.98 1.439 -13.188 <0.0001 Mean altitude 2.375*10-2 2.196*10-3 10.811 <0.0001 Mean slope angle 1.763 0.286 6.165 <0.0001 Mean slope angle2 -9.656*10-2 1.870*10-2 -5.165 <0.0001

88 Results

Figure 92. Relationship between sorted stripes and mean altitude and mean slope angle (for more details see Fig. 53).

Figure 93. Observed and predicted distribution of sorted stripes in the whole modelling area (for more details see Fig. 54).

89 Results

Sorted stripe distribution Non-sorted solifluction terrace distribution

18 7 16 14 6 12 5 10 4 8 3 6 4 2 2

1 Change in deviance (|x|) 0 0 Explained variance (%)

Peat cover Mean altitude Water cover Aspect majority Relative radiation Mean slope angle Mean shrub cover Rock terrain cover Mean canopy cover Elevation-relief ratio Std of wetness indexMean wetness indexConcave topography y-coordinate x-coordinate Sand and gravel cover MeanAutocovariate altitude Glacigenic deposit cover Aspect majority Relative radiation MeanMean slope angleshrub cover Rock terrain cover Std of wetness indexConcave topography Glacigenic deposit cover Figure 95. Results of the univariate analyses for non-sorted solifl uction terrace distribution (for Figure 94. Results of the hierarchical partitioning more details see Fig. 52). analyses for sorted stripe distribution (for more details see Fig. 55). tion in distribution of non-sorted solifl uction nation power than mean altitude, but still over terraces (Table 18). one percent. GLM and HP gave similar results In model evaluation, a Kappa measure of because altitude and slope angle were the most 0.047 (SE = 0.043, p = 0.204, cut off value = important predictors in both approaches. 0.26) indicated very poor discrimination ability of the model. The area under the curve of ROC plot (0.725, SE = 0.026, p < 0.0001) 5.3.9 Non-sorted solifl uction denoted a better prediction power than the κ terraces value, but it was still almost poor. The rather low prediction capability of the distribution Altogether 266 (11.7%) of the 2272 model- model and the difference between the Kappa ling squares were found to contain active non- and AUC values can also be seen in Figure sorted terraces. In univariate analyses, stand- 96. However, by means of the extrapolated ard deviation of topographical wetness index, calibration model, it was possible to discover mean slope angle, mean altitude, mean of top- a few new feature sites because four of the ographical wetness index and relative solar ra- eleven visited commission error squares were diation were the most important correlates for present sites. distribution (Fig. 95). The fi nal logistic regres- The autocovariate explained 10.5% and x- sion model included three predictors (Table coordinate 1.7% of the variance in the non- 30). The probability of feature occurrence in- sorted solifl uction data in hierarchical par- creased with increasing standard deviation of titioning (Fig. 97). Peat cover, mean altitude, topographical wetness index as well as mean mean slope angle and standard deviation of altitude and decreasing peat cover. All varia- topographical wetness index were the most bles were linearly associated with the response. important environmental factors but clearly The fi nal model explained 10.1% of the varia- less effective than the autocovariate. In the

Table 30. Variables, coeffi cients, standard errors (SE), z- and p-values of the fi nal model for non-sorted solifl uction terrace distribution. Variable Coeffi cient SE z p Intercept -8.588 1.049 -8.191 <0.0001 Std of wetness index 1.335 0.222 6.026 <0.0001 Mean altitude 8.734*10-3 1.593*10-3 5.484 <0.0001 Peat cover -2.352*10-3 4.480*10-4 -5.249 <0.0001

90 Results

Figure 96. Observed and predicted distribution of non-sorted solifl uction terraces in the whole modelling area (for more details see Fig. 54). end, the results of the GLM and HP were model with a positive sign (Table 31). Other fairly similar despite that slight differences oc- predictors in the model were aspect major- curred (Table 30 & Fig. 97). ity, mean of canopy cover and proportion of concave topography that all had a negative effect on the sheet occurrence. The affect of 5.3.10 Sorted solifl uction sheets concave topography on the feature probability was non-linear. In Figure 99 are displayed The number of active present feature squares the response curves of slope angle, altitude used in the modelling was 404 (17.8%). The and canopy cover. The fi nal model explained univariate analyses exhibited the overwhelm- 60.1% of the deviance change in the sheet ing effect of mean slope angle and mean al- data (Table 18). titude on sorted sheet occurrence (Fig. 98). The evaluation measures indicated very These topographical parameters were also the good and excellent discrimination ability of most important predictors in the fi nal logistic the logistic regression model. The Kappa

91 Results

Non-sorted solifluction terrace distribution Sorted solifluction sheet distribution

120 12 100 10 80 8 60 6 40 4 20 2

Change in deviance (|x|) 0 0 Explained variance (%)

Peat cover Mean altitude Water cover Peat cover Aspect majority x-coordinate y-coordinate Mean slope angle Mean shrub cover Relative radiation AutocovariateMean altitude Elevation-reliefMean ratio canopy cover Rock terrain cover Aspect majority Concave topographyMean wetness index Std of wetness index Mean slope angle Relative radiation Sand and gravel cover Elevation-relief ratio Glacigenic deposit cover Std of wetness index Concave topography Glacigenic deposit cover Figure 98. Results of the univariate analyses for Figure 97. Results of the hierarchical partitioning sorted soliﬂ uction sheet distribution (for more de- analyses for non-sorted soliﬂ uction terrace distri- tails see Fig. 52). bution (for more details see Fig. 55).

Table 31. Variables, coeffi cients, standard errors (SE), z- and p-values of the fi nal model for sorted solifl uction sheet distribution.

Variable Coeﬃ cient SE z p Intercept -7.887 2.585 -3.051 0.0023 Mean slope angle 0.934 6.753*10-2 13.833 <0.0001 Mean altitude 3.001*10-2 3.134*10-3 9.576 <0.0001 Aspect majority -5.557*10-3 1.106*10-3 -5.025 <0.0001 Mean canopy cover -0.103 2.515*10-2 -4.094 <0.0001 Elevation-relief ratio -6.037 1.500 -4.025 <0.0001 Concave topography -0.276 7.643*10-2 -3.608 0.0003 Concave topography2 2.675*10-3 7.472*10-4 3.580 0.0003 value was 0.636 (SE = 0.039, p < 0.0001) and high probability (> 0.8) were in the northeast AUC of ROC plot was 0.943 (SE = 0.009, p and east (Fig. 100). Eight of the commission < 0.0001). The best Kappa measure was ob- error sites were visited, but none of them were tained with the cut off value of 0.44. Most of found to contain sorted sheets. the present squares with low feature probabil- The results of the HP denoted the impor- ity (< 0.2) were in the margins of larger sheet tance of the mean altitude, mean slope angle, ﬁ elds, and almost all absent squares with very mean of shrub cover, mean of canopy cover

Figure 99. Relationship between sorted soliﬂ uction sheets and environmental variables (for more details see Fig. 53).

92 Results

Figure 100. Observed and predicted distribution of sorted solifl uction sheet in the whole modelling area (for more details see Fig. 54). and glacigenic deposition cover from the set (Fig. 102) but the proportion of explained of environmental predictors (Fig. 101). In ad- deviance decreased rapidly between the three dition, autocovariate variable explained 7.8% most effective predictors. The results of the and y-coordinate 1.1% of the variation in the generalized linear model showed that the cov- response data. Thus, the altitude and slope er of sorted sheets increased when mean alti- angle were the most important environmental tude as well as mean slope angle increased and predictors in both statistical approaches. relative radiation at fi rst decreased but then in- Abundance modelling of sorted sheets was creased (Table 32). The fi nal abundance model conducted with all 404 present squares. The explained ca. 46 % of the deviance change in cover of the feature ranged from 0.01 ha to the data (Table 18). 16.06 ha in the modelling squares. Mean alti- The assumption of normal errors was tude, mean slope angle, aspect majority and correct because residuals of the model were relative solar radiation had the greatest effect randomly distributed when they were plot- on the deviance change in univariate analyses ted against fi tted values. Furthermore, distri-

93 Results

Sorted solifluction sheet distribution mental variables explained over one percent 8 7 of the variation, namely mean altitude, mean 6 5 slope angle, standard deviation of topographi- 4 3 cal wetness index and mean of shrub cover. 2 1 Therefore, mean altitude and mean slope angle 0

Explained variance (%) were the most important determinants in both approaches but the effect of relative radiation

y-coordinate x-coordinate AutocovariateMean altitude decreased substantially in HP. Aspect majority MeanMean slope angleshrub cover Relative radiation Mean canopy coverConcaveElevation-relief topography ratio Glacigenic deposit cover 5.3.11 Sorted solifl uction streams Figure 101. Results of the hierarchical partitioning analyses for sorted solifl uction sheet distribution A total of 233 (10.3%) modelling squares were (for more details see Fig. 55). found to contain active sorted streams. Mean altitude, mean slope angle and elevation-relief Sorted solifluction sheet abundance ratio caused the largest changed in deviance in 25 univariate analyses (Fig. 104). The rest of the 20 variables affected clearly less than the above 15 mentioned. The variables and coeffi cients of 10

5 the ﬁ nal logistic regression model for feature

Change in deviance (|x|) 0 distribution are presented in Table 33. The probability of feature occurrence increased

Peat cover Mean altitude Water cover with increasing mean altitude and mean slope Aspect majority Mean slope angleRelativeMean radiation shrub cover Rock terrainElevation-relief cover ratio Mean canopy cover Std of wetnessMean wetness index index Concave topography Sand and gravel cover Glacigenic deposit cover Sorted solifluction sheet abundance

Figure 102. Results of the univariate analyses for 12 sorted solifl uction sheet abundance (for more de- 10 8 tails see Fig. 52). 6 4 2 bution of the residuals was close to normal. 0 The Spearman’s rank correlation coeffi cient Explained variance (%) for the calibration data was 0.650 (p < 0.0001) x-coordinate y-coordinate AutocovariateMean altitude Aspect majority Relative radiation Mean slope angleMean shrub cover Rock terrain cover and for the evaluation data 0.539 (p < 0.0001). Elevation-relief ratio Std of wetness index Thus, some level of degradation of prediction Glacigenic deposit cover power occurred when the calibration model was evaluated. Based on hierarchical partition- Figure 103. Results of the hierarchical partitioning ing, the autocovariate was the best explana- analyses for sorted solifl uction sheet abundance tory variable (Fig. 103). Four of the environ- (for more details see Fig. 55).

Table 32. Variables, coeffi cients, standard errors (SE), t- and p-values of the fi nal model for sorted solifl uction sheet abundance. Variable Coeffi cient SE t p Intercept 13.45 5.170 2.602 0.0098 Mean altitude 8.4287*10-3 6.693*10-4 12.593 <0.0001 Mean slope angle 0.103 1.976*10-2 5.198 <0.0001 Relative radiation -0.426 0.127 -3.358 0.0009 Relative radiation2 2.715*10-3 7.859*10-4 3.455 0.0006

94 Results angle. Response curves of mean altitude and sion error squares were found to be present mean slope are illustrated in Figure 105. The sites in the fi eld evaluation. fi nal model explained ca. 45% of the deviance According to the hierarchical partitioning, change in the response data (Table 18). the most effective environmental predictors Cohen’s Kappa statistic for evaluation data were mean altitude, mean slope angle, mean indicated good discrimination power of the of canopy cover, mean of shrub cover, mean model (κ = 0.565, SE = 0.052, p < 0.0001, cut of topographical wetness index and glacigenic off value = 0.31). The area under the curve deposit cover (Fig. 107). The mean altitude ex- measure was 0.951 (SE = 0.009, p < 0.0001) plained slightly more of the variation in the which indicated an excellent performance of data than autocovariate. In general, the results the model. The good prediction ability of the of the GLM and HP approaches were similar. model can also be seen from Figure 106. Most of the present sites with low predicted probability (< 0.1) were on the edges and absent 5.3.12 Defl ations sites with high probability (> 0.8) in the core distribution area. Five of the seven commis- A total of 290 (12.8%) of the 2272 modelling squares were found to contain defl ations. Ac- Sorted solifluction stream distribution cording to univariate analyses mean of canopy

60 cover, mean altitude, standard deviation of 50 topographical wetness index, sand and gravel 40

30 cover as well as mean of shrub cover were 20 the most important variables for the deﬂ a- 10 tion distribution (Fig. 108). It is notable that

Change in deviance (|x|) 0 the canopy cover and mean altitude changed the deviance much more than the other vari- Peat cover Mean altitude Water cover Aspect majority Mean slope angle RelativeMean radiation shrub cover ables. The results of the GLM are presented Elevation-relief ratio Rock terrain coverMean canopy cover Std of wetness index MeanConcave wetness indextopography Sand and gravel cover Glacigenic deposit cover in Table 34. The probability of fi nding defl a- tions in a modelling square increased when the Figure 104. Results of the univariate analyses for proportion of sand and gravel as well as mean sorted solifl uction stream distribution (for more de- altitude increased and mean of canopy and tails see Fig. 52).

Figure 105. Relationship between the sorted soliﬂ uction stream occurrence and mean altitude and mean slope angle (for more details see Fig. 53).

95 Results

Table 33. Variables, coeffi cients, standard errors (SE), z- and p-values of the fi nal model for sorted solifl uction stream distribution.

Variable Coeﬃ cient SE z p Intercept 16.24 -1.042 -15.59 <0.0001 Mean altitude 2.723*10-2 2.130*10-3 12.78 <0.0001 Mean slope angle 0.472 4.618*10-2 10.22 <0.0001

Figure 106. Observed and predicted distribution of sorted soliﬂ uction streams in the whole modelling area (for more details see Fig. 54). shrub cover decreased. The response curves model explained 20.5% of the deviance in the of sand and gravel and mean altitude were data (Table 18). humped expressing the non-linear association The evaluation measures indicated a fairly between the distribution of features and pre- good discrimination ability of the distribu- dictors. The shapes of the response curves are tion model. The Kappa value was 0.446 (SE = illustrated in Figure 109. The ﬁ nal distribution 0.053, p < 0.0001) and AUC of ROC plot was

96 Results

Sorted solifluction stream distribution ture probability (< 0.1) occurred on the fells

6 or fell slopes and most of the absent squares 5 with rather high probability (> 0.3) were in 4 3 the valleys. Nineteen of the commission er- 2 ror squares were visited, and thirteen of them 1 0 were found to contain defl ations. Explained variance (%) The autocovariate explained 11.6% of the variance in the response data that was over y-coordinatex-coordinate MeanAutocovariate altitude Aspect majority Mean slopeMean angle shrub cover Relative radiation two-fold more than the most important envi- Mean canopy cover Elevation-relief ratio Mean wetness index Glacigenic deposit cover ronmental variable, sand and gravel cover (Fig. 111). In addition, x-coordinate explained 1.6% Figure 107. Results of the hierarchical partitioning of the variance indicating a slight geographi- analyses for sorted solifl uction streams distribution cal trend in the response data. The other most (for more details see Fig. 55). infl uential environmental predictors were standard deviation of topographical wetness index, glacigenic deposit cover and canopy Deflation distribution cover. The results of the GLM and HP had 45 notable differences. Mean altitude and mean 40 35 of shrub cover were independently measured 30 25 clearly less important than standard deviation 20 15 10 of topographical wetness index and moraine 5 cover, which were excluded from the fi nal dis-

Change in deviance (|x|) 0 tribution model.

Peat cover Mean altitude Water cover Aspect majority Mean shrub coverRelative radiation Mean slope angle Mean canopy cover Elevation-relief ratio Rock terrain cover Std of wetness indexConcaveMean topography wetness index Sand and gravel cover Glacigenic deposit cover 5.3.13 Modelling results: a summary Figure 108. Results of the univariate analyses for deﬂ ation distribution (for more details see Fig. 52). The modelling results are summarised in Ta- 0.812 (SE = 0.028, p < 0.0001). The optimal bles 35 and 36, and Figures 112 and 113. In κ measure was obtained with cut off value of distribution modelling, the amount of ex- 0.31. The predictions calculated for the whole planed deviance varied from 7.6% (stony earth modelling area are presented in Figure 110. circles) to ca. 60% (sorted soliﬂ uction sheets). Many of the present squares with low fea- In abundance modelling, values ranged from

Table 34. Variables, coeffi cients, standard errors (SE), z- and p-values of the fi nal model for defl ation distribution.

Variable Coeﬃ cient SE z p Intercept -20.64 4.798 -4.301 <0.0001 Sand and gravel 3.772*10-3 4.489*10-4 8.404 <0.0001 Sand and gravel2 -1.056*10-6 2.281*10-7 -4.628 <0.0001 Mean canopy cover -9.865*10-2 1.559*10-2 -6.327 <0.0001 Mean altitude 0.109 2.527*10-2 4.294 <0.0001 Mean altitude2 -1.521*10-4 3.341*10-5 -4.554 <0.0001 Mean shrub cover -4.95*10-2 1.383*10-2 -3.577 0.0003

97 Results

Figure 109. Relationship between deﬂ ations and sandy soils and altitude (for more details see Fig. 53).

Figure 110. Observed and predicted distribution of deﬂ ations in the whole modelling area (for more details see Fig. 54).

98 Results

Deflation distribution ca. 37% (palsas) to ca. 52% (earth hummocks). 14 12 In model evaluation, the kappa values ranged 10 from 0.047 (non-sorted soliﬂ uction terraces) 8 6 to 0.641 (peat ponus) and AUC measures 4 2 from 0.722 (sorted nets) to 0.951 (sorted solif- 0

Explained variance (%) luction streams). In the abundance modelling, earth hummock and peat pounu models ob-

x-coordinate y-coordinate Autocovariate Mean altitude tained the best prediction ability (R > 0.7). Aspect majority s Mean shrub cover Mean canopy cover MeanConcave wetness index topography Sand andStd gravel of wetness cover index The number of the predictors varied from two Glacigenic deposit cover to eight (distribution model of peat pounus) in the ﬁ nal GLMs. In addition, models included Figure 111. Results of the hierarchical partitioning commonly at least one non-linear factor. analyses for deﬂ ation distribution (for more details see Fig. 55).

Table 35. Summary of the model performances of the distribution models. Kappa and area under the curve (AUC) values are for evaluation data [italic = good prediction ability, bold = excellent prediction ability (Landis & Koch 1977: 165; Swets 1988: 1292)]. Landform Change in Kappa AUC deviance (%) (eval) (eval) Palsa 52.3 0.565 0.950 Convex non-sorted circle 33.8 0.482 0.879 Stony earth circle 7.6 0.254 0.735 Earth hummock 47.0 0.637 0.920 Peat pounu 46.9 0.641 0.913 Stone pit 18.8 0.318 0.806 Sorted net 14.2 0.170 0.722 Sorted stripes 43.7 0.437 0.937 Non-sorted solifl uction terrace 10.1 0.047 0.725 Sorted solifl uction sheet 60.1 0.636 0.943 Sorted solifl uction streams 45.2 0.565 0.951 Defl ation 20.5 0.446 0.812

Table 36. Summary of the model performances of the abundance models. Spearman’s rank correlation coefﬁ cients (Rs) are for calibration (cal) and evaluation (eval) data.

Landform Change in Rs Rs deviance (%) (cal) (eval) Palsa 37.1 0.647 0.274 Convex non-sorted circle 39.0 0.633 0.650 Earth hummock 51.9 0.757 0.722 Peat pounu 47.2 0.689 0.715 Sorted soliﬂ uction sheet 45.6 0.650 0.539

99 Results

Figure 112. Summary of the environmental factors of the ﬁ nal distribution models. The direction of the effect is indicated with signs (+ = positive correlate, − = negative correlate, + − = non-linear correlate with a humped response curve, − + = non-linear correlate with a downward humped response curve).

Figure 113. Summary of the environmental factors of the ﬁ nal abundance models (for more details see Figure 112).

100

6 DISCUSSION

6.1 Periglacial landforms: 1997; Luoto & Hjort 2004). prevalence, distribution, The diversity of the periglacial landforms activity and morphology varied spatially although some clumped patterns could be identifi ed from the study area. In northernmost Finnish Lapland, periglacial Analysis squares without periglacial features phenomena are common close or above the were commonly characterised by fairly low treeline. However, the frequencies of the fea- slope angles and dry soils. However, the veg- tures and diversity of the types vary spatially etation types and densities as well as altitudes (e.g. Piirola 1969, 1972; Seppälä 1982b, 1987, can vary considerable between the non-occur- 1993b, 1997b; Kejonen 1997; Hjort & Seppälä rence sites. Squares with only one landform 2003; Luoto & Hjort 2004, 2005). In the study type were commonly occupied by the extensive area, the prevalence of the periglacial land- fi elds of earth hummocks, peat pounus, palsas forms varied from 0.2% to 65.4% at the 25-ha or inactive sorted nets. Considerably many of resolution. The most abundant features were the modelling squares included two to four dif- earth hummocks, peat pounus, sorted nets and ferent landform types, which is mainly caused sorted solifl uction sheets. The specifi c envi- by the presence of feature continuums (Figs. ronmental requirements and fairly continental 30 & 114; see also Karte 1979: 146; Ballantyne climate of the study region are probably the & Harris 1994: 205–209). The most common main reasons for the small number of slope sequences were: (1) earth hummock – stone pit features such as slushfl ow tracks, ploughing – peat pounu – palsa, (2) sorted stripe – sorted and braking blocks, non-sorted solifl uction solifl uction stream – sorted solifl uction sheet, lobes and taluses (e.g. Rapp 1986; Pissart 1993). (3) sorted polygon/circle – sorted net – sorted Moreover, the low activity of certain processes stripe and (4) non-sorted step – convex non- may cause the rarity of, for example, non-sort- sorted circle – earth hummock/active sorted ed and sorted polygons, snow accumulation net (cf. Piirola 1969; Seppälä 1982b; van Vliet- hollows and stone pavements. In general, the Lanoë & Seppälä 2002). The sites with highly prevalence differences between different land- variable environmental conditions often had form types were quite expected if compared more than four types of periglacial landforms. to other subactic regions (cf. Lundqvist 1962; Commonly, these squares included upper fell Harris C. 1982; Jahn & Siedlecki 1982; Meier areas with sorted patterned ground and slope 1987, 1991; Niessen et al. 1992). landforms as well as valleys with non-sorted The environmental conditions of the stud- circles and solifl uction features. ied area have promoted the formation of dif- The activity of periglacial landforms ferent cryoturbation and peat accumulation changes temporally as well as spatially and it based non-sorted features (e.g. van Vliet-Lanoë is fairly challenging to classify features to dis- and Seppälä 2002). For example, earth hum- crete active/inactive classes (e.g. Goldthwait mocks and peat pounus covered over 15% of 1976; Matthews et al. 1998; Haugland 2004; the study area. Also the sorted landforms were also see below). The activity classifi cation rather common periglacial features. For in- method used in this study was rather sugges- stance, the total cover of sorted nets and solif- tive and it probably emphasized the number luction sheets was over 24 km2 (4%). However, of the active landforms because the classifi - the abundance of the sorted features describes cation was only used to separate clearly inac- mainly the former conditions and periglacial tive landform occurrences from at least partly processes rather than the present circum- active features. Therefore, fi eld measurements stances in northernmost Finland (see Kejonen of the processes should be performed to draw

101 Discussion

Figure 114. Schematic cross-section summarising the study results. Proﬁ le (A): the main topographical, soil and vegetation characteristics of the study area. Proﬁ le (B): the occurrence of the modelled periglacial landform types and other geomorphological features. more detailed conclusions of the activity of et al. 2005). However, to utilise the modelling the periglacial features in the region (cf. Ke- approach presented in this study to explore jonen 1994; Seppälä 2005c). Nonetheless, the effects of climate change on phenomena based on the mapping results, it can be seen occurrence direct environmental determinants that northern Finnish Lapland is a transitional should be included in the set of predictors and zone between active continuous permafrost the causality of these factors should be studied regions and stabilized relict feature areas (cf. in more detail. Jahn & Siedlecki 1982). In general, the sensi- Palsas are clearly the most studied and most tive subarctic regions are key areas when ad- accurately described periglacial landforms in dressing the potential consequences of cli- northern Finland (e.g. Auer 1924, Ruuhijärvi mate change on periglacial phenomena (e.g. 1960, 1962; Salmi 1970, 1972, Seppälä 1976b, Zuidhoff 2002; Luoto et al. 2004b; Walker et 1983, 1986, 1994b, 2003a; Luoto & Seppälä al. 2004). For example, future environmental 2002a; Rönkkö & Seppälä 2003). In general, changes could affect the activity of landforms the palsas in the study area resembled those and processes in a few decades, especially on described in the literature (e.g. Åhman 1977; small-sized and permafrost-related features Seppälä 1988; Zuidhoff & Kolstrup 2005) al- (e.g. Ballantyne & Matthews 1982, 1983; van though some differences existed. For example, Vliet-Lanoë et al. 1993; Nelson 2003; Fronzek string-form palsas were smaller and had no

102 Discussion specifi c orientation compared to mire inclina- the distribution of the features. Hence, the tion (cf. Salmi 1972: 130; Åhman 1977: 41). In convex non-sorted circles were classifi ed into addition, small rather ambiguous humped fea- a separate group and a descriptive term was tures were distinguished from classical coni- applied. The new term was used to avoid the cal palsas into the class of hummocky palsas. confusion of these landforms with the other From the different morphological types (Åh- non-sorted circles. Nevertheless, the termino- man 1977: 38–42), palsa plateaus and distinct logical vagueness of the different non-sorted esker palsas could not be found in the study circles complicated the comparison between region despite that some large tortuous ridge- the features described in the literature and features were detected. circles found from the study area (e.g. Shilts The activity determination of the palsas 1978; Harris 1998; Overduin et al. 2003; Ping was rather problematic because most of the et al. 2003; Walker et al. 2004). In addition, the landforms are at mature stage and obviously morphological distinction of the convex non- degrading (see Figs. 3 & 33; cf. Ruuhijärvi sorted circles from other patterned ground 1962; Zuidhoff & Kolstrup 2000; Luoto et was occasionally diffi cult in the fi eld. For ex- al. 2004b). Moreover, the permanency of the ample, if sorting was present, the landform re- frozen ground found from some of the string- sembled poorly sorted nets (Fig. 35; cf. Moo- form and hummocky palsas was diffi cult to de- ers & Glaser 1989: 429; also see Walker et al. termine. The presence of permafrost should 2004: 172). be determined in the end of the thaw season The morphological and distributional char- (Seppälä 1976b, 1983) but this was not pos- acteristics of stony earth circles have not been sible due to the large research area. Further- studied in Finnish Lapland. In the study area, more, the core may be frozen for several years the stony earth circles resembled convex non- or decades but thaw during a rainy summer sorted circles, but features were (1) on average (Seppälä, personal communication). Thus, the smaller, (2) fl at without raised centres and (3) distinction between real palsas and large peat had scattered appearance and formed rarely pounus can be diffi cult without several year- continuous pattern fi elds (Williams 1959; Riss- round measurements (e.g. Lundqvist 1969: ing & Thorn 1985; Wilson & Sellier 1995; cf. 213–214). In addition, Seppälä (1998) has ar- Chapter 5.2.2). Nonetheless, some continuous gued that some of the peat pounus are per- fi elds were found, for example from northeast mafrost features. Nevertheless, all palsas were of Suohpášoaivi fell and east of Guovdoaivi defi ned to be active and indicators of sporadic fell. permafrost and thereby fulfi lled the activity Earth, turf and peat hummocks are often requirement for the modelling (cf. Lundqvist combined in Finnish literature (e.g. Salmi 1972; 1962: 73). van Vliet-Lanoë & Seppälä 2002; Seppälä In Finland, studies on non-sorted circles 2005c). However, based on the fi eld observa- have focused on earth and peat hummocks tions, it was presumable that slightly different (e.g. Salmi 1972; Seppälä 1998, 2005c; Luoto environmental factors determine the distribu- & Seppälä 2002b; van Vliet-Lanoë & Seppälä tion of mineral and peat hummocks (see also 2002) and observations on non-sorted circles Lundqvist 1969: 210–211). Thus, the mineral of other kinds are rather limited (Seppälä soil cored features and pure peat hummocks 1982b, 1987). The convex non-sorted circles were classifi ed to separate groups. Both clas- described in this study resembled stony earth sical earth hummocks (e.g. Schunke & Zoltai circles (Chapter 5.2.3) and earth hummocks 1988) and thin peat mounds (Raup 1965; van (Chapter 5.2.4) but had a characteristic appear- Vliet-Lanoë &Seppälä 2002) were combined ance (cf. Washburn 1979: 128-129; Ballantyne under the earth hummock term for practical & Harris 1994: 91). Due to the morphological reasons (Lundqvist 1969: 210–211; Niessen et and distributional differences, it was presum- al. 1992: 194; cf. Grab 2005: 140). ably that slightly different processes governed Due to the relatively continuous vegeta-

103 Discussion tion cover, the activity determination of the (1962: 21–24) and differed clearly from Wash- earth hummocks was in several circumstances burn’s (1979: 129, Fig. 5.13) stone pits. rather challenging. For example, large pattern Sorted patterned ground features are rather fi elds were located in the forested areas where common in northern Finland (e.g. Piirola 1969, the snowcover can be rather thick, over one 1972; Seppälä 1982b, 1987; Tabuchi & Hara metre, which reduces the intensity of frost 1992; Hirvas et al. 2005; Luoto & Hjort 2005). processes (e.g. Clark et al. 1985: 213). How- However, most of the observations have been ever, using quite detailed observations of the made of sorted circles and polygons, or at least vegetation cover, frost heaving and general soil the features have been termed so (see Chapter disturbance the activity of the earth hummock 2.2.3; Lundqvist 1962: 29; Ballantyne & Harris areas, even in the mountain birch forests, was 1994: 195). In this study, if the patterns were determined (see also Jahn & Siedlecki 1982: not clearly circular or polygonal the feature 43; van Vliet-Lanoë & Seppälä 2002: 187; Sep- was determined to be sorted net (Washburn pälä 2005c: 175). 1956: 830). Therefore, most of the landforms Peat pounus are probably the most studied on gentle slopes with elongated mesh patterns patterned ground in Finnish Lapland (Ru- were classifi ed to the net and not to the circle uhijärvi 1960: 220–222; Salmi 1972; Seppälä or polygon class. In general, the morphology 1998; Luoto & Seppälä 2002b; Seppälä 2005c). and size of the nets vary considerably in the Morphologically, peat pounus are rather simi- study area (cf. Washburn 1969: 163; Ballantyne lar to earth hummocks (Chapter 5.2.4; van & Matthews 1982: 345; Coxon 1988: 83–84). Vliet-Lanoë & Seppälä 2002). However, in this The mapping of the activity of the sorted study, the pounus were separated from earth nets was a rather straightforward task. How- hummocks because of the material difference; ever, the abundance of the nets and the re- peat pounus are formed of organic material activation of the old landforms may have led and they lack of mineral core (Lundqvist 1969: to some misinterpretations. The oldest sorted 212; cf. van Vliet-Lanoë & Seppälä 2002). Fur- nets were probably formed in severe climate thermore, the large peat pounus (> 70 cm conditions close to the retreating ice edge dur- high) often have a perennially frozen core (e.g. ing the deglaciation (Luoto & Hjort 2004: 334; Salmi 1972: 124; Luoto & Seppälä 2002b: 129), Chapter 3.6; cf. Cook-Talbot 1991: 138; Mat- whereas permafrost in mineral soil features is thews et al. 1998: 161–162; Haugland 2004: rather exceptional in Finnish Lapland (Seppälä 299). Nowadays, these inactive pattern fi elds 1997b: 91–93). The activity determination are more abundant than active feature areas. of the features was slightly more straightfor- Thus, despite that the sorted nets are very ward than for the earth hummocks because common landforms, the cover of the active the surfi cial characteristics of the active and nets did not vary enough between the mod- inactive pounus differed more clearly. In addi- elling squares for abundance modelling. For tion, it was possible to use the observations of example, over 60% of the present modelling the frozen cores in the classifi cation (see also squares had active features less than 0.1 ha and van Vliet-Lanoë & Seppälä 2002: 187; Seppälä only four squares had landform cover more 2005c: 175). than one hectare. Stone pits are small sorted circles that have Sorted stripes are fairly abundant patterned mainly been described from Sweden (e.g. ground in Finnish Lapland (Piirola 1969, Lundqvist 1962: 21–24). The features have 1972; Seppälä 1972b; Kejonen 1979; Söder- been mentioned in Finnish literature (e.g. Aar- man 1980; Hirvas et al. 2005). In the study tolahti 1969: 17; Piirola 1972: 133; Seppälä area, the landforms resembled sorted streams 1982b: 237, 1987: 48), but their environmental and had rather similar distribution in the land- properties have not been studied in detail. In scape (Appendices 8 & 11). In addition, the general, the stone pits mapped from the study formative processes of stripes, frost heaving area resembled those described by Lundqvist of stones from ground moraine, solifl uction

104 Discussion and eluviation of fi nes (see Washburn 1956: some misinterpretations probably exist in the 857–859, 1969: 185–188), are also possible data. mechanisms for the sorted streams genesis Sorted solifl uction sheets have been rec- (cf. Clapperton 1975: 212–216). Despite the ognised from northern Finnish Lapland, but morphological and distributional similarities they have often been termed as block slopes feature types were separated because stripes (Piirola 1969: 19, 1972: 133; Kejonen 1979: belongs to the patterned ground and streams 9–10). In this study, all solifl uction formed are considered to be solifl uction landforms block and boulder slopes, ancient or present, (e.g. Washburn 1947: 87). The activity classi- were included into the class of sorted solifl uc- fi cation of the sorted stripes was rather chal- tion sheets (cf. Ballantyne & Harris 1994: 209– lenging. Most of the stripes were probably 210). The transverse nature of the landforms formed during or shortly after the deglaciation in the landscape is probably due to large-scale but have reactivated under colder and moister step-like slope topography that is caused by conditions after climatic optimum (Kejonen bedrock schistosity (Mikkola 1932: 32–33). 1994: 54; Chapter 3.6; cf. Ballantyne 1984: These large-scale risers are generally steeper 314). However, if the feature, recently formed than adjacent steps. or ancient, had some sign of current activity, it The activity determination of the sorted was classifi ed into the active class. sheets was rather troublesome because of the Hult (1887) and Helaakoski (1912) were features’ age. All sorted sheets were most prob- probably the fi rst who described solifl uction ably formed before ca. 8600 BP and stabilized phenomena in Finland. Afterwards, proc- during the milder phase when forests covered esses and landforms have been studied for the area (Eronen 1979; Kultti 2004). After the example by Ohlson (1964), Piirola (1969), climatic optimum ca. 4000 BP, many of the Kejonen (1979), Seppälä (1979b), Söderman sheets have reactivated and are active under (1980), van Vliet-Lanoë et al. (1993) and Mat- the present climatic conditions (Kejonen 1994; thews et al. (2005). The non-sorted solifl uction cf. Ballantyne 1984: 325). Moreover, the activ- terraces have been described from different ity determination was in some places compli- parts of Finnish Lapland (e.g. Kejonen 1979: cated by two facts. Firstly, the whole sheet can 10; Hirvas et al. 2005) but their distributional be in a uniform movement despite the stones characteristics have not been studied in de- and blocks being covered by lichen (active but tail. The features in the Báišduattar – Áilegas seem to be inactive). Secondly, lichen cover can area resemble those described in the literature be incomplete because of snow accumulation (e.g. Söderman 1980: 126; Ballantyne & Har- (nivation) processes and wind defl ation (inac- ris 1994: 205–206; cf. Lundqvist 1962: 51–53). tive but seem to be active). Both of these facts However, the terraces were generally rather can give a false impression of the activity of small with frontal raiser less than 50 cm high the landform. However, detailed fi eld observa- (cf. Jahn & Siedlecki 1982: 36–37). Due to the tions were used to ensure the correctness of landforms small size, only their locations were the classifi cation. In addition, the fi eld meas- mapped and distribution modelled. The non- urements of Kejonen (1994) and the appear- sorted terraces were divided into active and in- ance of frontal soil embankments (e.g. Fig. 48) active based on the observations of vegetation gave further confi dence about the activity or characteristics and morphology of the fea- inactivity of a specifi c sheet (see also Jahn & tures (e.g. Ballantyne & Harris 1994: 213). In Siedlecki 1982: 36). addition, the overall picture of the activity of Studies on sorted solifl uction streams solifl uction features and their appearance was are quite limited despite the fact that sorted obtained from the studies of Kejonen (1979), slope features are rather common in Finnish Söderman (1980) as well as Jahn and Siedlecki Lapland (Piirola 1969, 1972; Seppälä 1972b, (1982). Still, the activity determination of the 1982b; Kejonen 1979, 1994; Söderman 1980). non-sorted terraces was rather diffi cult and In general, the material and possible formative

105 Discussion processes of the sorted streams resembled the was very low when the mean slope angle was material and probable formation mechanisms over ca. three degrees, because palsas occurred of the sorted stripes and sheets in the study on fl at areas (Lundqvist 1969: 208), in moist area (cf. Clapperton 1975: 212–214). Further- and fl at-bottomed valleys that have commonly more, there was a morphological continuum been bottoms of paleolakes, i.e. ice-dammed between the sorted streams and largest stripes lakes (see Chapter 3.3). Furthermore, the slope (e.g. Washburn 1969: Fig. 114). Therefore, the variable is comparable to the proportion of earlier discussion on the origin and activity of fl at topography (see Luoto & Seppälä 2002a: sorted solifl uction landforms also holds here 23) that was excluded from the fi nal set of pre-

(see also Tyurin 1983: 1284). dictors based on the correlation analysis (Rs of Defl ations can be found from different mean slope angle and proportion of fl at to- parts of Finnish Lapland where late glacial pography was −0.849). and early Holocene sand dunes are abundant The inclusion of water cover with a (e.g. Tanner 1915; Ohlson 1964; Seppälä 1971, humped response curve was partly contradic- 1974, 1982, 1993a, 1995b; Tikkanen & Heik- tory because a water body as an energy store kinen 1995; Käyhkö et al. 1999; Kotilainen is a negative factor for palsa preservation 2004). In addition to the sandy soil areas and (e.g. Matthews et al. 1997: 120; Sollid & Sør- moraine hummocks, wind has also eroded bel 1998: 288). However, Luoto and Seppälä peat deposits on the fell summits (e.g. Seppälä (2002a) also derived a water variable in their 1972b; Luoto & Seppälä 2000) and palsa mires distribution model of palsas. They explained (e.g. Seppälä 2003a) in the study area. the positive association of palsas and water bodies with the exposure of open mire areas to strong winds and moisture availability (Luoto 6.2 Environmental factors & Seppälä 2002a: 25). However, the presence affecting periglacial landform of water variable can also indicate the occur- occurrence rence of thermokarst ponds that are abundant in large palsa mires, especially in palsa complexes (cf. Luoto & Seppälä 2003). 6.2.1 Palsas Univariate analyses emphasized that shrub cover and rock terrain affect the palsa distri- In Báišduattar – Áilegas, palsas were common bution. Despite that shrub cover was not in on level peaty areas where the small ponds the fi nal logistic regression model, it could were abundant. However, if the water cover have an important negative effect on the palsa exceeded a certain threshold in the 25-ha grid formation because dense shrub cover enables cell, ca. 5 ha when the water variable was fi tted considerable snow accumulation. The thick separately, the probability of the feature occur- snow layer would avert the deep frost forma- rence decreased displaying a humped relation- tion and thus prevent permafrost growth (e.g. ship. Not surprisingly, peat cover was the most Goodrich 1982; Seppälä 1982a, 1994b). How- important factor in the distribution model of ever, if the palsa has formed, the denser shrub palsas (cf. Luoto & Seppälä 2002a) because the cover may not cause the degradation of the palsa formation and preservation is dependent palsa (cf. Rönkkö & Seppälä 2003: 999; Sep- on the organic material that enables frost for- pälä 2003b; Zuidfoff & Kolstrup 2005: 58). mation during the winter and protects the per- The importance of rock terrain is slightly pe- mafrost core from thawing during the summer culiar. Naturally, there occurs a negative corre- (e.g. Brown 1970: 176; Åhman 1977: 129–130; lation between palsa distribution and rock ter- Seppälä 1988: 268–269; Zuidhoff & Kolstrup rain but this association was rather weak and 2005: 49). it was emphasized due to the utilised univari- The slope angle was a strong negative cor- ate approach in this study (see Crawley 1993: relate for the palsa occurrence. The probability 192).

106 Discussion

The model included three geomorphologi- for cold air, soil moisture and snow. Therefore, cally realistic factors that explained over half the effect of this topographical curvature vari- of the deviance change in the data. Thus, the able is contradictory. distribution model for palsas was quite robust The abundance of palsas was determined despite that the response was rather strongly by two already discussed factors, namely peat spatially autocorrelated. In addition, the fi eld and shrub cover. Instead, the difference be- checking revealed that the model was even tween the results of the fi nal abundance model better than expressed by the evaluation meas- and the univariate analyses needs further nota- ures (cf. Bustamante & Seoane 2004). Howev- tion. The minor importance of the peat in the er, despite the good assessment results, there univariate approach is probably attributed to occurred prediction errors that can be caused the correlation of the variable with the shrub by six main reasons. cover, mean altitude and slope angle (Table First, the peat cover of the site was too thin 14). However, the slope variable was not in for palsa formation. Second, the mire is too the abundance model because peat and shrub wet for palsa preservation. Third, palsas could cover are probably suppressing it, i.e. the pre- form on a relative small mire, if the peat layer dictors are strongly intercorrelated. The nega- was thick enough and the mire was open to the tive association of palsas with altitude can be wind effect. Fourth, the set of the used predic- explained by climatological and biological factor variables was not complete, if we take into tors. During the winters, the valleys and lower account all the important causal factors for altitudes can be colder due to temperature palsa formation and preservation (e.g. Seppälä inversion and cold air drainage (Harris S.A. 1986: 141–145, 1988: 268–270). The solution 1982b; Fig 28C; Chapter 3.6). Moreover, the for the previous problems could be the utilisa- drier soil conditions at higher altitudes are un- tion of more causal correlates, for example a suitable for peat and palsa formation (e.g. Sep- variable describing the thickness of the peat pälä 1986: 142; cf. Luoto & Seppälä 2002a: 25). layer. In general, the problems concerning The association between the palsa abundance the use of direct determinants such as snow- and elevation-relief ratio, as well as standard cover and temperature in spatial palsa models deviation of topographical wetness index, has are discussed in Luoto and Seppälä (2002a: to be viewed at a 25-ha grid resolution. The 25–26). Fifth, despite that the palsa data was high values of moisture variability and the low based on detailed photo interpretation as well values of elevation-relief ratio describe con- as extensive fi eld mapping, there can be minor cave valleys that are potential sites for peat ac- defi ciencies in the response data (see Chapter cumulation and palsa formation. 6.1; cf. Brown 1970: 178). Sixth, the grid ap- The fi nal abundance model included two proach can cause some problems in the detec- realistic predictors that explained rather well tion of the key variables (see Chapter 6.3.3). the deviance change of the response. Still, the Hierarchical partitioning showed that the prediction ability of the abundance model was soil moisture, shrub cover, altitude and con- relatively poor because the calibration model cave topography were potentially important could not predict the palsa cover satisfacto- variables that were not in the fi nal distribu- rily with the evaluation data. The failure can tion model. The association between palsas be attributed to two reasons: (1) the model- and shrub cover was already discussed and ling square selection or (2) a lack of important the effect of altitude is treated below. The causal factors from the fi nal model. In the fi rst topographical wetness index as an indicator case, the selection procedure neglected mod- of the soil moisture and location of streams elling squares with substantial palsa cover be- could be an important positive or negative fac- cause these squares also had abundant water tor for the palsas (Brown 1970: 176; Seppälä cover. By removing these hot spots of palsas 1986: 142; Luoto & Seppälä 2002a: 20, 2003: valuable information was lost for the model 25). The concave areas are accumulation sites building. However, the criterion for omitting

107 Discussion abundant water cover squares was justifi ed in for the formation of non-sorted circle (e.g. the modelling process (cf. Chapter 4.1.2). In Washburn 1969: 111; Schunke & Zoltai 1988: the second case, the inclusion of variables 235, 239–240). The concave areas are also ac- representing soil moisture, slope angle, wa- cumulation sites for wind-drifted snow, which ter cover and concave topography should be have a negative effect on the soil freezing, re- explored in more detail because the presence spectively. However, the raised centres of the of spatial autocorrelation indicates that some circles enable deeper frost formation into the causal factor is missing from the model (see features than into the vegetated inter-circle Chapter 6.3). boundary sustaining the soil particle motion The differences between the distribution (cf. Overduin et al. 2003: 870–872; Walker et and abundance models were expected because al. 2004: 182). they address different aspects of the same phe- The altitude had a clear non-linear effect on nomena. In distribution modelling, the whole the probability of landform occurrence, and study area was utilised in the analyses, squares this concorded with the observations made in with and without palsas, whereas the abun- the fi eld. The convex non-sorted circles oc- dance modelling was focused on the present curred at high altitudes above the treeline but squares only. The peat cover was the only still in the valleys, which are favourable sites common but also the most important corre- for frost processes, for example cryoturbation, late for palsa distribution and abundance. The to act (e.g. van Vliet-Lanoë 1988a, 1991). In mean slope angle was not in the abundance addition, the direct solar radiation was dem- model because present squares had rather onstrated to be an important discriminator uniform topography, whereas slope angle was between the occupied and unoccupied mod- an important discriminator in the distribution elling squares. The radiation differences may modelling. The same holds for the water cover cause variability in the soil moisture and veg- as well but the exclusion of the water variable etation density that can initiate circle genesis from the abundance model can also be due to (e.g. Shunke & Zoltai 1988: 240; Walker et al. the procedure of modelling squares selection. 2004: 183–185). The shrub cover was not a proper discrimina- According to the univariate analyses, the tor but it was an important factor in the pre- canopy cover and topographical wetness index diction of palsa abundance. were potentially important correlates of the convex non-sorted circle distribution. Both of the variables were realistic factors, as stated 6.2.2 Convex non-sorted circles earlier. However, these variables were not in the fi nal GLM model probably because of the In the study area, convex non-sorted circles intercorrelation of explanatory variables; the were common in rather fl at-bottomed val- canopy cover correlates with altitude (Rs = leys at relatively high altitudes (cf. Hodgson −0.729) and the wetness index with concave

& Young 2001: 320–323). To be more precise, topography (Rs = 0.633). the circles were often situated in large cirque- The variables in the fi nal distribution model like bedrock valleys that are relatively com- were rather plausible albeit strong spatial au- mon in the study region (Kaitanen 1969: 46). tocorrelation and trend was present in the re- Based on literature (e.g. Washburn 1969: 108, sponse data. In addition, based on the evalu- 118; Overduin et al. 2003: 869), non-sorted cir- ation measures, the distribution model was cles occur on level ground and, therefore, the good in discriminating between the occupied negative effect of the slope angle on the prob- and unoccupied sites. Nonetheless, there oc- ability of the feature occurrence was expected. curred prediction errors which were mainly Moreover, the concave sites are accumulation caused by (1) the lack of important causal areas for a fi ne-grained soil material and mois- factors from the fi nal model, (2) defi ciencies ture that are, in general, crucial determinants in the mapping and (3) due to the grid ap-

108 Discussion proach (see Chapter 6.3.3). Firstly, the utility 1965: 417–418) and it could be a meaningful of vegetation and soil variables could improve correlate. The standard deviation of wetness the model performance (see also below). For index describes valleys that are potential sites example, the glaciofl uvial activity could have for the convex non-sorted circles (see above). washed the fi nes away from the glasigenic de- The set of the most important environ- posits generating unsuitable soils for the non- mental factors for the circle distribution and sorted circles to form (Shunke & Zoltai 1988: abundance was strikingly similar. The surro- 235). Furthermore, many of the topographi- gates for topography, climate and soil mois- cally proper sites were covered by peaty soil or ture were present in both GLM models. The rather dense shrub cover, factors that prevent only signifi cant difference between the models the circle formation. Secondly, the fact that ca. was the effect of canopy cover. The tree cover 35% of the absent sites with high predicted could be a meaningful discriminator between probability were actually present squares indi- the present and absent sites but it does not cated the possibility to build better model if have an effect on the cover of the convex non- all, even the smallest feature fi elds, could be sorted circles because the feature fi elds occur observed. The large convex non-sorted circle mainly in the treeless regions. areas are detectable from aerial photographs but the smallest fi elds, less than 30 m in diameter, were diffi cult to interpret by remote sens- 6.2.3 Stony earth circles ing data and the discovery of the least active landforms was troublesome even in the fi eld. Based on the literature, stony earth circles com- Hierarchical partitioning analysis revealed monly occur in exposed, snow-free sites in the the potential importance of the canopy cover subarctic and on the alpine tundra where the and wetness index. These variables were not in vegetation cover is sparse (e.g. Williams 1959: the fi nal distribution model but their impor- 13; Williams & Smith 1989: 163–164; cf. Ping et tance was highlighted by the univariate analy- al. 2003). According to the distribution model ses as well. The canopy cover and soil moisture presented in this study, the stony earth circles could be more causal environmental correlates occurred in dry and sparsely vegetated moraine for the presence of convex non-sorted circles soils at relatively high altitudes. Hence, despite than the mean altitude and proportion of con- that the prediction performance and explana- cave topography (cf. Hodgson & Young 2001: tion power of the model were relatively poor, 321–323). the environmental determinants in the fi nal The abundance model was able to predict GLM model concord well with the previously the cover of the convex non-sorted circles made observations. fairly well despite that the response data was More precisely, the stony earth circles were spatially autocorrelated and had a strong trend common in open mountain birch forests and toward northeast. In addition, according to the on the alpine heaths where the undergrowth evaluation, the model was robust (cf. Chapter was relatively scant (cf. Rissing & Thorn 1985: 6.2.1). The most important landscape scale 154). The defoliated and poorly recovered factors were the altitude, soil moisture and birch forest region in the southeast was one of slope angle (e.g. Washburn 1956: 841–842; the core distribution areas of the stony earth Overduin et al. 2003: 870–872). The same circles (cf. Chapter 3.8). The relationship be- variables were the most infl uential also in the tween the circles and tree cover was also illus- univariate analyses. Moreover, the proportion trated by the non-linear effect of altitude on of concave topography, slope direction and the feature occurrence. The highest probabil- moisture variability could potentially affect the ity was at the treelimit, at ca. 400 m a.s.l. In this circle abundance. The slope direction as a sur- zone, the sparse vegetation cover enables de- rogate for solar radiation could refl ect the local fl ation and accelerates short-term freeze-thaw climatological differences of a site (e.g. Geiger cycles (Williams 1959: 3, 5; Rissing & Thorn

109 Discussion

1985: 154; Josefsson 1988: 220). fi ne-grained soil areas where shrubs were rela- The sites of stony earth circles were, based tively common (cf. Williams 1959: 13; Wash- on the topographical location, relatively dry. burn 1969: 106; Salmi 1972: 127; Tarnocai & However, the glacigenic deposits with wide Zoltai 1978: 591–592; Schunke & Zoltai 1988: grain-size distribution and relatively good 235). The peat and shrub cover had a non- moisture holding capacity often fulfi lled the linear connection to the feature distribution moisture requirement for the frost activity (e.g. Fig. 71). At fi rst, peat has a positive ef- (Goldthwait 1976: 30). In addition to the vari- fect on the mound occurrence because of the ables already in the distribution model, uni- material supply and moisture holding capacity variate analyses pointed out the possible im- (Raup 1965: 55; Niessen et al. 1992: 194; Grab portance of the solar radiation and proportion 2005: 143). However, in the end, the areas of of concave topography. Despite that these extensive peat accumulations were unsuitable predictors represent the vegetation and soil sites for the earth hummocks and were occu- moisture distribution in the area neither of the pied by peat pounus (Chapter 6.2.5) and palsas variables has, in general, a direct connection to (Chapter 6.2.1). the feature distribution (e.g. Williams 1959). The presence of shrubs, commonly dwarf The prediction ability of the model was birch (Betula nana), in the feature fi elds could be rather poor and several reasons for the failure a cause or an effect. The shrubs induce uneven can be proposed. The stony earth circles are snow distribution that is seen to be an impor- quite small and sporadically distributed non- tant factor in the earth hummock genesis (e.g. sorted patterned ground (cf. Overduin et al. Schunke & Zoltai 1988: 239). On the other 2003: 869). The scattered appearance impeded hand, the small-scale topographical irregular- their mapping and complicated the building ity caused by hummocks can be the reason for of a comprehensive response data set for the the shrubs presence because the features give analyses. In addition, the very strong spatial the required shelter for vegetation growth. In autocorrelation indicated a lack of causal fac- addition, the earth hummocks and shrubs may tors from the fi nal distribution model. For develop side by side. The mounds give shelter example, hierarchical partitioning analyses re- from strong winds and shrubs ensure mois- vealed a moderate independent importance of ture availability by collecting snow (see Kojima the shrub cover variable. In the end, the scale 1994: 266). Whatever the truth, too dense a difference between the formative mechanisms shrub cover has a negative effect on the proc- and utilised environmental variables (Williams ess activity and occurrence of the earth hum- 1959: 12–13) and unpredictable factors, such mocks (e.g. Lundqvist 1969: 211). as vegetation destruction by reindeer herding The negative association between hum- and direct human activities (Lundqvist 1962: mocks and altitude can be attributed to two 28; Seppälä 1987: 49; Harris 1998: 443), may main reasons (cf. Grab 2005: 150). Firstly, cause precariousness in the modelling. There- the lower altitudes are accumulation areas for fore, the stony earth circles may not be suited moisture, fi ne-grained soil material and cold for landscape scale distribution modelling, at air (cf. Chapter 6.2.2). Secondly, many of the least in a predictive manner. However, the rela- low altitude valleys with abundant earth hum- tive realistic model output indicated the pos- mock cover have been bottoms of paleolakes, sibility to use stony earth circles in explorative i.e. ice-dammed lakes during the deglaciation analyses. stage (Appendix 4: van Vliet-Lanoë & Seppälä 2002: 189–191). These silty soils are very frost susceptible and thus ideal areas for the forma- 6.2.4 Earth hummocks tion of earth hummocks (Shilts 1978; Tarnocai & Zoltai 1978: 592; Schunke & Zoltai 1988: Based on the distribution model, active earth 239; van Vliet-Lanoë 1988a: 92). hummocks occurred often in moist peaty and The variables in the fi nal distribution mod-

110 Discussion el covered relatively well the set of important soil data could even further improve the pre- landscape scale factors for the earth hum- diction ability of the model (Schunke & Zoltai mock occurrence albeit the response data was 1988: 239). spatially autocorrelated (cf. Schunke & Zoltai 1988; van Vliet-Lanoë & Seppälä 2002). More- over, the model was able to explain almost half 6.2.5 Peat pounus of the deviance change in the hummock data. Therefore, the very good prediction ability of Based on the logistic regression model present- the model was expected. However, the inclu- ed in this study, the active peat pounus were sion of soil factors could even further improve common in moist relatively fl at-bottomed val- the prediction performance of the model. leys where the peat and shrub cover was high The HP analysis revealed the possible im- (cf. Salmi 1972: 124–129; van Vliet-Lanoë & portance of concave topography and slope Seppälä 2002: 189–190). Moreover, the feature angle. The proportion of concave topography fi elds gained generally less solar radiation and was positively (cf. Grab 1997: 443–444) and were more often covered by mountain birch mean slope angle negatively associated with forests than unoccupied sites. Smaller peat the earth hummock distribution in the earlier pounu occurrences were relatively abundant stages of the model building, but these vari- in the cirque-like valleys but at lower altitudes ables were omitted from the fi nal model prob- than the convex non-sorted circles (Chapter ably due to the intercorrelation. For instance, 6.2.2) and earth hummocks (Chapter 6.2.4). slope is one of the factors controlling the The fi nal model for the peat pounus in- earth hummock distribution in the landscape cluded eight environmental predictors. The (van Vliet-Lanoë & Seppälä 2002: 189) but large array of correlates can be attributed to the explanation power of slope was taken by the intermediate location of the peat pounus the wetness index, peat cover as well as shrub in the landscape, spatially between the earth cover (see Table 14). hummocks and palsas, and heterogeneity of The abundance of earth hummocks was the environmental conditions of the present determined by the soil properties and topo- sites (cf. van Vliet-Lanoë & Seppälä 2002: graphical factors. Hence, rather the same set 190). The largest mires and thickest peat de- of predictors had an effect on the feature posits were occupied by palsas and most of distribution and abundance. However, some the mire edges were covered by mineral-cored minor differences existed between the models. mounds. Therefore, the distribution of pou- The abundance model included the slope an- nus could not be explicitly determined by the gle and water cover variables and the excluded utilised set of predictors, at least not at the 25- standard deviation of topographical wetness ha modelling resolution. However, despite that index factor. For example, the inclusion of the model had a high number of predictors, the slope was understandable because exten- almost all of the variables were explainable sive earth hummock fi elds occur in fl at areas with distinct physical connection to the pounu (Mackay & Mackay 1976: 889; Jahn & Siedlecki distribution. 1982: 41). Organic material is a basic requirement for The fi nal abundance model was rather peat pounu formation (e.g. Salmi 1972: 124; good, regardless of the presence of spatial Seppälä 1998) and the humped response curve autocorrelaton, because the model explained is due to the presence of palsas in extensive over 50% of the deviance change in the re- mires (Fig. 77). Flatness, concavity, shadiness, sponse data and it predicted the earth hum- moderate vegetation cover and soil moisture mock cover fairly well. In addition, the model are all positive factors for peat accumulation was relatively robust despite that six explana- and pounu development (van Vliet-Lanoë & tory variables were present (cf. Crawley 1993: Seppälä 2002: 187–190). For example, concav- 211). However, the inclusion of high quality ity promotes cold air ponding and deep frost

111 Discussion formation in the pounu hummocks (Seppälä sided valleys and in the breaks on otherwise 1998: 369–371; cf. Chapter 6.2.2). At fi rst, the relatively steep slopes (cf. van Vliet-Lanoë & water cover can be seen as a negligible predic- Seppälä 2002: 189–190). tor but small ponds are often indicators of The differences between the results of the mires and relatively high ground water level (cf. abundance model and univariate analyses were van Vliet-Lanoë & Seppälä 2002: 190). Thus, considerable. The slope angle was important only the inclusion of the canopy cover into in both approaches but the peat cover and the fi nal model was rather diffi cult to explain elevation-relief ratio were distinctively less because pounus had no causal connection to meaningful in the univariate analyses. The lat- the mountain birch forests (see also Luoto & ter approach indicated connection between Seppälä 2002: 129). However, the pounus were landform abundance and water cover, sand often located at relatively low elevations and and gravel, soil moisture, rock terrain, concave the altitude variable was rather highly corre- topography and canopy cover variables. In lated with the canopy cover (see Table 14). theory, sand and gravel as well as rock cover The distribution model for peat pounus was are negatively correlated to the pounu abun- quite robust because even the Kappa measure dance but the association was rather weak in indicated almost excellent prediction perform- the study area (cf. Chapter 6.2.1). The rest of ance, and the model was able to explain over the predictors may have a stronger relation to 46% of the deviance change. In addition, based the response despite that the variables were on the fi eld evaluation, the model could predict not in the fi nal model. In addition, hierarchi- the present squares correctly despite that the cal partitioning displayed the potential impor- pounus were not previously observed because tance of the shrub cover (van Vliet-Lanoë & all visited high probability absent squares were Seppälä 2002: 187–190). actually present sites. However, even though Rather clear differences between the vari- the model had a high ability to distinguish be- ables in the fi nal distribution and abundance tween occupied and unoccupied squares, some models occurred (see Chapter 6.2.1). The only prediction errors occurred as well. Most of shared predictors were peat cover and slope the incorrectly predicted squares were present angle, which were important both in discrimi- sites with low predicted probability (omission nating between occupied and unoccupied sites error) (see Chapters 4.2.3 & 6.3.3). Altogether, but also in determining the feature cover. The to build even better distribution models for shrub, water and tree cover, proportion of pounus, the number of the predictors should concave topography, standard deviation of be reduced (e.g. Crawley 1993: 188, 211) and wetness index and direct solar radiation were the role of the wetness index as well as mean proper discriminators, whereas elevation-re- altitude explored. lief ratio compiled the set of important cover The prediction ability and explanation predictors. power of the abundance model were good and the included predictors were quite realistic despite the presence of autocorrelation in the 6.2.6 Stone pits response data. Peat cover had a positive, and slope angle as well as elevation-relief ratio a Active stone pits were often found from moist negative, effect on the pounu cover. However, and fl at valley fl oors where the ground was when the effect of slope was examined sepa- covered by shrubs and glacigenic deposits rately, the upward curvature of the response were fairly abundant. The slope angle was the curve of the slope variable increased when most important landscape scale factor with a the slope angle surpassed a certain threshold negative effect on the feature occurrence (e.g. (ca. seven degrees). The shape of the curve Lundqvist 1962: 22). The importance of fl at can probably be attributed to the presence of topography was also highlighted by the shape small pounu fi elds in the bottoms of steep- of the response curve. The predicted prob-

112 Discussion ability decreased drastically when slope angle feature formation. Fourth, the presence of increased (Fig. 84). The shrub cover had an spatial autocorrelation and minor trend in the indirect connection to stone pit distribution response data indicated that some important because features occurred in areas of abun- determinants could be missing from the fi nal dant shrub cover despite that the stone pits distribution model. For example, the soil mois- does not require dense vegetation to form (e.g. ture indicators, such as topographical wetness Lundqvist 1949: 336). index, proportion of concave topography and Topographically the features were found peat cover might improve the discrimination from valleys where the ground water level was power of the model (see above). close to the ground surface (Aartolahti 1969: 17). The presence of ground water and its fl uctuations enabled wash out of fi nes from 6.2.7 Sorted nets the frost-sorted stony pits (Seppälä 1987: 48). Addition to the requirements for the stone In the Báišduattar – Áilegas area, the active pit formation, the soil has to be rather rich in sorted nets were more common in valleys at stones and boulders (e.g. Lundqvist 1949: 336; relatively high altitudes where soil moisture see Chapter 3.3; cf. Kessler & Werner 2003: was present. Based on the logistic regression Fig. 3). However, the squares with the high- model, the standard deviation of topographi- est probabilities were those where the moraine cal wetness index was the most important pre- cover was at the intermediate values. This is dictor for feature distribution. In general, this explained by the presence of peat in the fea- variable cannot be seen as the most impor- ture squares. The peat does not have direct tant correlate for sorted net occurrence (e.g. connection to the stone pits but features oc- Washburn 1969: 164). However, the valleys are cur in thin peat areas, on the edges of mires potential sites for nets, if other environmen- (see Seppälä 1987: 48). In addition, according tal conditions are favourable for frost sort- to univariate analyses, soil moisture and sorted ing processes (see e.g. Corte 1966: 237; Kling material could be potential variables in distri- 1998: 444; Matsuoka et al. 2003: 75). bution modelling. The association of stone Indicator of soil moisture and surrogate for pits and soil moisture is understandable, as climate conditions were in the model but one discussed above, but the connection to sand crucial determinant, namely soil material (i.e. and gravel is without a clear foundation. moraine cover), was not in the fi nal model (cf. The discrimination ability of the model Goldthwait: 1976: 30). Instead, the peat cover was quite good, if we take into account the was included into the distribution model with fi eld checking results where almost 70% of the a humped response curve (Fig. 88). Nonethe- absent squares with relatively high predicted less, based on the fi eld observations, a logical probability were present sites. However, as explanation for this exists. The net fi elds were the evaluation measures indicated, numerous often located close to the mires in thin peat incorrectly predicted squares existed. Based areas where soil moisture was abundant and on the visual analyses of the prediction map proper material for sorting was present (Wash- and fi eld observations, four main reasons for burn 1969: 164; Hodgson & Young 2001: the errors could be proposed. First, the small 320). size and the fairly scattered distribution of the The factors in the distribution model were stone pits caused defi ciencies in the response mostly explainable but still the prediction abil- data (cf. Aartolahti 1969: 17), but this holds ity of the model was, at the most, moderate. only for the unvisited sites. Second, the soil It is diffi cult to fi nd unambiguous causes for material was in some sites probably too dry the prediction errors but four potential rea- and the ground water level too deep in the sons are proposed. First and probably the ground for the landform genesis. Third, small- most important reason could be the scale dif- scale topographical variability may prevent the ference between the causal factors and utilised

113 Discussion environmental variables. For example, the the average slope angle was over ca. 10º the importance of small-scale determinants for site was occupied by sorted solifl uction sheets patterned ground can be seen from the study (Chapter 6.2.10), streams (Chapter 6.2.11) or of Matthews et al. (1998). The second poten- talus formations (Goldthwait 1976: Fig. 4; Bal- tial reason for prediction errors could be the lantyne 1984: Fig. 6). Besides the slope gradi- heterogeneity of the sorted nets and forma- ent, the altitude is a good determinant of slope tive processes (Washburn 1970: 440–441). activity in Finnish Lapland (Seppälä 1993b: Morphology of the patterns can vary notably 60). The higher altitudes, generally areas above between the different net fi elds and some of 400 m a.s.l., are treeless heaths with low veg- them can be still evolving toward sorted circles etation cover, which enables solifl uction proc- or polygons (e.g. Washburn 1969: 165; Ballan- esses to act (e.g. Benedict 1976: 60). tyne & Matthews 1982: 344–345; Ballantyne Altitude and slope angle were the most im- & Harris 1994: 93). This may cause problems portant factors also in the univariate analyses. in the modelling because, according to Kling The other notable predictors in the univariate (1998), sorted circles and polygons occur in analyses, namely canopy cover, slope aspect, different locations in the landscape and the concave topography and shrub cover were all primary formative factors are different. How- possible and realistic correlates despite that ever, based on the fi eld observations, this was they were excluded from the fi nal GLM model a minor problem in the Báišduattar – Áilegas (Washburn 1969: 179; Pissart 1993: 212–213; area and the fi ndings of Kling (1998) may not Matsuoka 2001: 127, 130). In the earlier steps hold in all regions and every circumstance. of the model building, the mean of canopy The third reason for the errors could be the cover, aspect majority, proportion of concave small size of the active pattern fi elds that led topography and mean of shrub cover were all to mapping diffi culties (cf. Kling 1998: 444). negative factors for the occurrence of stripes. Finally, the presence of spatial autocorrelation Based on the predictors above, sorted stripes and trend in the sorted net data indicated a can be found from rather sparsely vegetated lack of direct factors from the fi nal model. For convex slopes, which are often orientated to example, the concave topography, slope angle the eastern sector (cf. Ballantyne 1984: Fig. and till cover could be more causal correlates 7). Nowadays, the slope aspect has a minor (Washburn 1969: 164, 167, 173; Goldthwait importance to the genesis of the features but 1976: 30; van Vliet-Lanoë 1988b: 1009–1012; when the stripes were formed, strong north- Kling 1998: 450; Matsuoka et al. 2003: 75). western winds transported snow (i.e. soil However, the fi nal distribution model was bet- moisture) to eastern and south-eastern slopes ter than indicated by evaluation measures be- (Seppälä 1993b: 272). However, strong winds, cause new sorted net sites were found with the especially during winters, blowing from the extrapolated calibration model. northwest may affect the distribution of active sorted stripes in the study area at present (cf. Hall 1994: 122; Holness 2001: 80). The 6.2.8 Sorted stripes formation of new stripes on the fell slopes is mostly prevented by the lack of soil moisture Most of the active sorted stripes occurred on and due to too dense vegetation cover (e.g. relatively steep slopes above 440 m a.s.l. in the Åkerman 1993: 247). Therefore, the recently study area. The inclusion of slope variable into formed features can only be found from spe- the model was axiomatic but the non-linear na- cial locations, such as moist valley fl oors (Fig. ture was not. Based on the fi eld measurements 46). However, these stripes cannot be detected and literature (e.g. Washburn 1969: 181; Wal- with the built distribution model because local ters 1983: 1351; Hall 1994: 121), sorted stripes small-scale factors dominate their formation do not occur on very gentle slopes with gradi- (e.g. Muir 1983; Werner & Hallet 1993). ents less than ca. three degrees. Equally, when The distribution model for sorted stripes

114 Discussion was robust. Firstly, the model could with only altitude is connected to process activity (Sep- two factors explain a notable amount of the pälä 1993b: 60) and the standard deviation of deviance change and, secondly, accurately wetness index describes topographically the discriminate between the occupied and unoc- potential occurrence areas (Ballantyne and cupied sites regardless of the spatial autocor- Harris 1994: 115, 205–206). relation and slight trend in the response data. The absence of causal correlates led to However, several prediction errors could also poor explanation power and prediction ability be identifi ed. More than half of the incorrectly of the model (see Fig. 96). On the other hand, predicted squares were absent sites with high soil moisture variability, altitude and peat cover predicted probability. Most of the commis- were among the four most important variables sion error squares were non-occurrence sites in the hierarchical partitioning where the ef- without any specifi c reason but some of the fect of spatial autocorrelation and geographi- squares were occupied by other sorted solif- cal trend was controlled by including spatial luction features. Moreover, some of the stripe variables. This denotes the distinct association occurrences were not previously detected in between the features and included predictors the fi eld. The omission errors were mainly at- at a landscape scale despite the fact that vari- tributed to the simplicity of the model; alti- ables are not direct determinants of the oc- tude and slope angle could not encompass all currence of solifl uction terraces (e.g. Benedict the potential factors determining the occur- 1976: 61–63; Matsuoka 2001: 127–128). In rence of stripes (Washburn 1956: 857–859). general, the inclusion of more direct factors For example, many landforms on the slopes such as slope angle and soil moisture could im- of large glacial melt water channels remained prove the model performance (e.g. Washburn unidentifi ed by the distribution model. 1999: 175). However, the wetness index is not Based on the fi eld observations and HP a suitable surrogate for soil moisture on slopes. analysis, the inclusion of topographical factors Instead, information of the perennial or late such as the standard deviation of topographi- laying snow patches could help to model the cal wetness index and proportion of convex distribution of the non-sorted solifl uction topography should be considered. In addi- terraces. As noticed in the fi eld and by many tion, the vegetation and soil variables could researchers, snow distribution, especially late improve the prediction ability of the model laying snow patches, are spatially connected to (e.g. Lundqvist 1962: 55; Washburn 1969: 179; solifl uction phenomena (e.g. Williams 1959: 5; Benedict 1976: 61–62). However, the robust- Ulfstedt 1993: 220). ness and simplicity of the constructed model In spite of all, the main reason for the mod- indicated good extrapolation potential of the el failure can probably be attributed to the size model (cf. Crawley 1993: 188, 211). and scattered occurrence of the features. Non- sorted solifl uction terraces are small periglacial features, which are distributed fairly randomly 6.2.9 Non-sorted solifl uction at a landscape scale. Nevertheless, the occur- terraces rences of the features were strictly controlled by topography on a local scale in the study According to the distribution model, the non- area. For instance, based on the fi eld obser- sorted terraces were frequently found from vations, glaciofl uvial melt water channels were valleys at high altitudes where the peat cover the most common occurrence sites. Therefore, was poor or absent. If we take into account the to improve the prediction power of the model important determinants for solifl uction fea- more specifi c topographical correlates should tures (e.g. Washburn 1979: 198–204; Åkerman be added to the set of predictors. For exam- 1993: 247; Pissart 1993: 212–213; Matsuoka ple, variables describing the length or number 2005: 41), all three variables in the model were of the melt water channels and foot slope ar- rather defi cient factors despite the fact that the eas should be considered. The effect of local

115 Discussion topography exceeded even the infl uence of one predictor was without causal connection soil material because the non-sorted solifl uc- to the response (see above). The good prediction terraces were as common in poorly sorted tion performance is seen to be caused by the sand and gravel areas as in moraine soils (cf. strong association between the sorted sheets Harris et al. 1995; Chapter 3.3). and slope angle as well as altitude. However, despite that the model was an excellent discriminator, it could not be use to detect new 6.2.10 Sorted solifl uction sheets landform sites. This is explained with the good accuracy of the response data, which is due to Based on the distribution model, active sorted large size of the sheets. Therefore, fi ve other sheets were common at high altitudes, on con- reasons are proposed for the causes of predic- vex southeast and northeast slopes (cf. Chapter tion errors. 6.2.8). When the slope angle and altitude were First, the topography of the site was unsuit- related separately to the occurrence of sheets, able for the sheets. For example, small-scale the most abrupt rise in the feature probability variability caused by the melt water channels occurred when the mean slope angle increased broke the unity of the slope, or the slopes from fi ve to 13º (cf. Ballantyne 1984: Fig. 6; were too gentle. The second reason, inability Åkerman 1993: 231) and mean altitude from of the utilised predictors to capture the ef- 400 m to 530 m a.s.l. The most important pre- fect of small-scaled topographical variability, dictors, slope angle and altitude, were rather is related to the previous one. Third, small self-evident landscape scale correlates (Piirola frost creep features that are not true sorted 1969: 32; Francou & Bertran 1997: 382; Mat- solifl uction sheets (cf. Kejonen 1979: 11) oc- suoka 2001: 127–128) although they have been cupied some of the high probability sites. effective in different periods. Slope gradient Fourth, some of the predicted sites were areas was the important factor when the sheets were of bedrock outcrops with thin moraine cover. formed, whereas altitude determines the dis- In some places, the sheet material can origi- tribution of the active landforms nowadays nally be from bedrock but more often thin soil (Seppälä 1993b: 60). sites are unsuitable for solifl uction phenom- Topographically the fell slopes can be di- ena (e.g. Åkerman 1993: 247). The potential vided into two main sections; upper convex importance of the soil material emerged also slopes and lower concave slopes. The convex in the hierarchical partitioning. The fi fth and slopes are areas for accelerated soil motion fi nal reason was the grid approach (see Chap- and thus the most probable places for sorted ter 6.3.3). Moreover, to improve the prediction solifl uction landforms, as indicated by the dis- ability of the distribution model, the role of tribution model (cf. Karte 1979: 146; Ballan- shrub cover should be explored because some tyne 1984: Fig. 7, Fig. 9; Ballantyne & Harris of the absent high altitude fairly steep slopes 1994: 205). Instead, concave slopes are sites of were densely vegetated. The association be- decelerated material movement and thus more tween the sorted sheets and shrub cover was probable areas for non-sorted slope features also shown in the HP analysis where the shrub (e.g. Washburn 1947: 87; cf. Chapter 6.2.9). variable was the most important predictor af- The inclusion of the elevation-relief ratio ter altitude and slope angle. Also, the canopy variable with a negative sign is more diffi cult cover could be a potential discriminator even to justify. However, this predictor may com- though it correlates rather strongly with alti- plete the fi nal set of determinants and act as a tude (Table 14). proper discriminator for some of the model- The cover of sorted sheets was determined ling squares without true causality. by altitude, slope angle and relative solar ra- The distribution model was robust albeit diation. In the abundance model, the altitude the data was clearly spatially autocorrelated was more important than slope angle. This and the model included six variables, of which is because the active fi elds are more exten-

116 Discussion sive at higher fell areas than at lower altitudes 6.2.11 Sorted solifl uction streams with corresponding slope gradients due to more favourable conditions for frost activity In the study area, active sorted solifl uction (see above). According to the solar radiation streams were common on the fell slopes at variable, sorted sheets were abundant on the high altitudes (cf. Piirola 1969: 17, 1972: 133; northern and southern slopes. The correlation Clapperton 1975: 212). The importance of the of the sheets with northern slopes was un- altitude and slope angle on sorted solifl uction derstandable because these slopes receive less phenomena was already discussed in Chapters solar radiation and remain colder, moister and 6.2.8 and 6.2.10 and only one additional com- consequently more active. The occurrence of ment can be made concerning the difference the sorted solifl uction sheets on the southern between the distribution of sorted streams and slope sector was more diffi cult to explain, al- stripes. According to Figure 105, the steepest though the association could be attributed to rise in the probability of sorted stream pres- the moisture distribution during the genesis of ence was when altitude was over 410 m a.s.l. the landforms (see Chapter 6.2.8). and slope angle was more than seven degrees. The abundance model was demonstrated to Thus, despite that the general distribution of be quite robust in the model evaluation. How- the sorted streams resembled stripes occur- ever, the prediction ability of the model was rence, the response curves and fi eld observa- better with building (Rs = 0.650) than testing tions indicated a minor distributional differ- data (Rs = 0.539). This rather substantial de- ence between these two landform types. The crease could be due to the lack of some causal stripes were more common at slightly higher factor from the model, which was also indi- altitudes on convex slopes where the mean cated by the presence of clear spatial autocor- gradient was less than on the middle parts of relation in the response data. According to HP, the slopes where the streams were more abun- the standard deviation of topographical wet- dant (Clapperton 1975: 212; Fig. 114). ness index and mean of shrub cover variables The variables in the fi nal distribution model explained independently part of the variance covered the most important landscape scale in the sheet data. From these predictors, the factors for the occurrence of sorted solifl uc- inclusion of the vegetation variable should be tion streams despite the fact that spatial auto- considered before the topographical surrogate correlation was present in the response data (e.g. Pissart 1993: 212; Ulfstedt 1993: 223). (cf. Tyurin 1983: 1283; Harris et al. 1998: 127). Based on the GLMs, a similar set of cor- Therefore, the good prediction ability of the relates determined the distribution and abun- model was expected. However, there still oc- dance of the sheets because altitude and slope curred prediction errors that can be explained angle were the most important factors in with four causes. Firstly, all features were not both models. In addition, the solar radiation detected in the fi eld because fi ve new occur- variable is comparable to slope direction albeit rences were found with the prediction map. that the predictors do not correlate strongly. Secondly, the topographical or soil properties However, the distribution model included of the site have been unsuitable for the feature three other variables which were not in the genesis. For example, glaciofl uvial melt water fi nal abundance model. The canopy cover, el- channels broke the unity of the slope or the evation-relief ratio and proportion of concave area was almost totally covered by bedrock. topography were not proper determinants for Thirdly, the site was occupied by some other sheet cover because landform fi elds occurred landform, usually sorted solifl uction sheet. in treeless areas where these topographical Finally, the local circumstances have enabled conditions were quite coherent. the formation of the features despite that the mean environmental conditions, according to the model, were unfavourable. Most of the sources of the prediction er-

117 Discussion rors could be removed with more accurate 253; Käyhkö et al. 1999: 435). These hum- data (e.g. Chapter 6.3.2) or by including new mocky landforms were usually very sparsely explanatory variables into the model (Harris et vegetated and were, therefore, the sites with al. 1998: 127). Hence, the effect of the canopy, the greatest potential for defl ation. shrub and moraine cover as well as soil mois- The effect of reindeer has to be taken into ture should be studied in more detail because account when the causes of defl ation are the active sorted streams occurred in fairly dry discussed (Seppälä 1984; Käyhkö & Pellikka and sparsely vegetated moraine soils above the 1994). As seen in the fi eld, reindeer were often treeline. On the other hand, the distribution grazing in the windy places where the mosqui- model with only two predictors explained a toes cause less harm. Consequently, the animals substantial amount of the deviance change in trample the vegetation and can infl ict local soil the data and the insertion of a variable that erosion leaving small bare soil patches, which would only slightly increase the explanation can be enlarged by wind action. Thus, some power of the model is unnecessary (e.g. Craw- of the defl ation surfaces were the result of the ley 1993: 188, 211). combined effect of wind and water erosion, as well as reindeer herding (Åkerman 1980: 253; Seppälä 1995b: 808–809). 6.2.12 Defl ations The univariate analyses and distribution modelling gave fairly similar results. However, According to the distribution model, defl a- the sand and gravel variable was less important tions occurred most often at relatively high than the standard deviation of topographical altitudes in sand and gravel soils where the wetness index in the univariate analyses. The vegetation cover was moderate or absent (cf. inclusion of the later predictor into the model Seppälä 1971: 22–29, 1984: 39, 2004b: 145; could be explained by the topographical loca- Åkerman 1980: 111; Käyhkö et al. 1999: 435). tion of the defl ations, but as indirect factor; it Defl ations were very common on glaciofl uvial would clearly be a less important correlate than and aeolian landforms where the sparse veg- the former variable. In addition, the standard etation cover could not resist the wind erosion deviation of wetness index variable correlates (Seppälä 2004b: 146). However, the non-linear moderately with the mean altitude as well as connection between the features and sandy the sand and gravel cover variables (Table 14). soils has to be due to the modelling resolution. The discrimination ability of the model Most of the present squares were mostly but was rather good, if we take into account the not totally covered with sand and gravel soil fi eld checking results where ca. 68% of the because of the linear appearance of the sorted absent squares with relatively high predicted deposits. Therefore, the response curve was probability were present sites. This defi ciency humped with the highest probability of fea- in the response data clearly demonstrated the ture occurrence at rather intermediate values. diffi culty in mapping small and quite scattered The negative association between the can- defl ations, although most of the features over opy cover and defl ation was explained by the fi ve metres in diameter were detectable from exposure of the open and sparse birch forest the aerial photographs. However, some other areas to stronger winds than densely forested causes for the predictions errors could also be regions (Seppälä 1993a: 274). However, the identifi ed. Firstly, many of the squares with humped response curve of mean altitude indi- abundant sand and gravel cover were without cated that the most probable sites were not the small-scale topographical variability, i.e. hum- fell summits despite that the strongest winds mocks and ridges, or the sites were densely and lowest vegetation cover occurred there. vegetated. For example, most of the esker Instead, the valleys with eskers, sand dunes, ridges had a thick tree cover that effi ciently palsas, kame- and moraine hummocks were prevented defl ation (cf. Seppälä 2004b: 146). the sites of wind erosion (e.g. Åkerman 1980: Secondly, the strong spatial autocorrelation in-

118 Discussion dicated that some important explanatory vari- data that may cause harm in statistical analyses able was missing from the model. HP dem- (e.g. Sokal & Rohlf 1981: 459–460; McCul- onstrated the independent importance of the lagh & Nelder 1989: 21–22). Furthermore, the standard deviation of topographical wetness causal factors of the studied phenomena may index and moraine cover. However, neither of operate on a different scale than the utilised these variables could be seen as a good addi- covariates (e.g. Luoto et al. 2001). Consequent- tional discriminator between occupied and un- ly, these data related problems and methodo- occupied defl ation sites. Perhaps the lack of a logical constrains should be taken into account predictor representing the hummocky topog- at the different stages of modelling in order to raphy could be one of the main reasons for obtain reliable results and to make valid inter- the errors. For instance, the defl ation surfaces pretations. were often present on the moraine landforms and the model could not predict these occurrences correctly. Moreover, the grid approach 6.3.1 Periglacial landform data may have introduced an additional source of error (Chapter 6.3.3). The utilised periglacial landform database is rather unusual because of its high mapping intensity, fi ne-scale resolution, and because it 6.3 Data and methodological consists of fairly precisely georeferenced sam- issues – advantages and pling locales. Furthermore, the same geomor- shortcomings phologist performed all surveys so variation resulting from differences between observers In general, numerical techniques and GIS does not exist. However, the response data data have not been widely used to model geo- also included inaccuracies and defi ciencies. morphological phenomena although a need The utilised digital aerial photographs were for cost-effi cient mapping approaches has ortho-rectifi ed. Still, there occurred spatial dis- emerged recenty (e.g. Bocco et al. 2001; Gude et continuity between neighbouring images. An al. 2002; Walsh et al. 2003b; Gurney & Bartsch additional inaccuracy was caused by the GPS 2005). In this study, the employment of a spa- positioning although the location errors are tial grid system with generalized linear model- estimated to be usually less than 20 m in open ling (GLM) and hierarchical partitioning (HP) mountain birch forests and better than 10 m proved to be a useful way to estimate the role in treeless alpine heaths (cf. Miettinen 2002: of several potentially signifi cant environmen- 43–54). Despite that these spatial inaccuracies tal predictors in determining the distribution do not generally cause serious problems in and abundance of periglacial landforms. How- mesoscale analyses, they should be considered ever, the utilisation of GIS data, a grid-based in the selection of the modelling resolution. approach and statistical methods in the mod- According to the fi eld checking, rather elling of geographical phenomena included many new previously unobserved landform several problems (e.g. Luoto & Hjort in press). sites were detected with extrapolated predic- Accuracy (e.g. thematic and positional), com- tion maps (e.g. Chapter 5.3.5), which is at the pleteness (areal coverage) and price are general same time a positive and negative issue. The GIS data problems (e.g. Burrough & McDon- new landform occurrences clearly demon- nel 2000: 220–221; Zhang & Goodchild 2002: strated the great potential of the numerical 3–6). In addition, statistical techniques com- approach to map periglacial features (cf. Bus- monly have several data related requirements tamante & Seoane 2004). On the other hand, that can be violated. For example, non-nor- new observations revealed defi ciencies in the mality, non-constant variance structure, spatial response data set. The most signifi cant prob- autocorrelation and intercorrelation of predic- lems occurred with small landform types (e.g. tors are general features of the geographical stony earth circles, stone pits etc.) that were dif-

119 Discussion

fi cult to map directly with remote sensing data spatial variables were only used in hierarchical and their observation was problematic even in partitioning to detect the independent contri- the fi eld (e.g. Chapter 5.3.6; Kling 1998: 444). bution of a specifi c predictor in this study (see In addition, all occurrences of the prevalent Chapters 4.2.3 & 6.3.3). landform types, such as earth hummocks, The selection of a link function for the re- were diffi cult to determine because of their sponses in the abundance modelling was prob- wide size and distribution range. Additional lematic due to the highly skewed distributions. problems were caused by activity (see Chapter Several different functions were tested but 6.1) and landform type defi nition. Periglacial none of them gave fair results. The test model- features often form continuums and it can ling showed that normalisation of the respons- be diffi cult to classify large landform fi elds to es could produce an appropriate outcome. discrete areas. For example, some of the cryo- Therefore, despite that the transformation can genic features have a rather similar morpho- complicate the results’ interpretation, the aim logical appearance (e.g. earth hummocks and was to normalise the responses. The utilised peat pounus) or complex genetic processes Box-Cox transformation defi nes the optimal (e.g. sorted nets). Therefore, it is quite obvi- transformation (e.g. Dobson 2002: 109), but ous that the large response database included still the responses were not ideally normally some misinterpretations. distributed, which can cause vagueness in the Spatial autocorrelation is a very general models. An additional possible problem in the statistical property of geomorphological vari- abundance modelling was caused by the clas- ables observed across geographic space (e.g. sifi cation system of the landforms because, Legendre 1993). Phenomena or environmen- in some cases, the landform fi elds mapped as tal variables are spatially autocorrelated if a points (0.04 or 0.16 ha) were detectable from measure made at one point can be predicted the residual plots (e.g. Fig. 64). However, it was with a measure made at another location, and assumed that this did not, in general, bias the autocorrelation is positive when subjects close modelling results. to each other are more alike than distant things The prediction abilities of the distribution (Goodchild 1986: 3). Spatial autocorrelation and abundance models (GLMs) were evalu- can hamper attempts to identify plausible re- ated by data obtained from the split-sample lationships between geomorphological phe- approach (e.g. Guisan & Hofer 2003). Conse- nomena and environment correlates, because quently, the data used in the model testing were the use of statistical tests may be invalidated spatially autocorrelated with the data used to by a strong spatial structure (e.g. Diniz-Filho et build the models and it is possible to gain over- al. 2003). For example, regression techniques optimistic estimates for the predictive abilities assume that modelled events are independent, of the models. Therefore, the utilisation of in- which is not true in the case of autocorrelated dependent evaluation data would improve the data (e.g. McCullagh & Nelder 1989: 21). reliability of the model predictions. In general, there occur several possible means to manage spatially autocorrelated data (Guisan & Zimmermann 2000: 156). For ex- 6.3.2 Predictor data ample, we can simply add an autocovariate to the set of predictors, widen the sample distance Key factors affecting periglacial landforms and beyond the distance of spatial autocorrelation processes occurrence are fairly well known or we can build an autoregressive model (e.g. (e.g. Washburn 1979: 10–17). However, to in- Legendre 1993; Augustin et al. 1996; Lichstein clude these determinants in the spatial model- et al. 2002). However, the introduction of an ling is problematic due to three main reasons. autocovariate can drop out some of the mean- Firstly, to generate spatial data layers of causal ingful predictor(s) because environmental correlates we often need numerous fi eld meas- factors are often also autocorrelated. Hence, urements, which are commonly laborious to

120 Discussion conduct. Spatially continuous information tant than absolute accuracy in the spatial analy- layers can be obtained by remote sensing but ses and the relative accuracy of the DEM was these techniques also pose some resolution, rather good (cf. Oksanen & Jaakkola 2000: 9). accuracy and price related disadvantages (e.g. Topography is probably the most widely Jensen 2000; cf. Klein 2004). Secondly, we can used surrogate of frost activity. A number of only perform point measurements of spatially studies have related different topographical variable environmental factors in the fi eld. parameters to periglacial landforms and proc- Thus, to build a continuous information lay- esses (e.g. Harris C. 1982; Ødegård et al. 1988; ers, we have to use interpolation techniques. Kling 1996; Matthews et al. 1998; Luoto & The resulting predictors are models that in- Hjort 2004, 2005). Not surprisingly, the mean clude a generally unknown amount of un- altitude-variable was selected in many models certainty. However, methods such as kriging, as a statistical signifi cant predictor. However, where the uncertainty can be presented may while the use of altitude variables in studies of help to spatially assess the validity of the data patterns of geomorphological features is wide- (e.g. Cressie 1993). Thirdly, the data used in the spread, it is also problematic. The altitudinal creation of a spatial information layer can be gradient represents surrogates for several en- inappropriate for the purpose. For example, vironmental gradients that are often intercor- the climatological variables interpolated based related, making tests of hypothesis associated on measurements conducted at the standard with these factors diffi cult and controversial. screen heights may not describe the actual For example, altitude has traditionally been conditions at the ground surface level where viewed as a surrogate for air temperature be- the processes act. Moreover, the network of cause temperature decreases 0.56–0.60°C/100 the meteorological stations is scattered in re- m (e.g. Geiger 1965). However, it should be mote regions and the measurements may only noted that the ‘classic’ elevation gradient of represent the climatological conditions of spe- periglacial features is obviously not caused cifi c topographical locations, for example val- by elevation per se. The elevation-dependence leys in mountainous regions (see Chapter 3.6; is ultimately caused by more causal edaphic Seppälä & Hassinen 1997: 158). Because of factors such as frost intensity, soil properties the lack of empirical data, mesoscale approach and vegetation (Matthews et al. 1998). In other and the issues discussed above, most of the words, elevation is a surrogate for one or more environmental factors used in this study were factors that are more direct correlates to geo- indirect correlates of periglacial phenomena morphologic processes than elevation. (cf. Guisan & Zimmermann 2000: 155). Soil moisture has been stated as one of the Several of the predictors were calculat- most important factors determining the dis- ed from the digital elevation model (DEM) tribution and activity of periglacial landforms (Chapter 4.1.3). In general, DEMs frequently (e.g. Ballantyne & Matthews 1982; van Vliet- contain systematic and non-systematic errors Lanoë 1988a; Matthews et al. 1998). The direct that are amplifi ed when fi rst (e.g. slope angle) factor representing moisture distribution could and second-order (e.g. wetness index) deriva- not be used in the modelling because empirical tives are calculated (Zhang & Goodchild 2002: information was not available. Hence, the top- 93–94, 120). Therefore, the quality of the cre- ographical wetness index was used in the anal- ated DEMs has to be assessed (Florinsky 1998: yses. Despite the indirect nature of the factor 41–42; Etzelmüller 2000: 140). The absolute and possible DEM data related uncertainties, height accuracy of the created DEM could not the wetness index is a greatly acknowledged be evaluated because detailed height measure- and widely used surrogate of soil moisture ments were not available from the study area. distribution (e.g. Moore et al. 1991). The im- However, the lack of reference information portance of the topographical wetness index does not cause serious problems because the was also highlighted in several distribution and relative accuracy of the DEM is more impor- abundance models in this study. However, the

121 Discussion spatial information of the springs could even misinterpretations that were probably be- further improve the possibilities to model the cause of the diffi culty to map areas with thin distribution of periglacial landforms on the soil cover. Thirdly, sand and gravel areas were fell slopes and foot of slopes because features, grouped together although the frost suscepti- for example earth and peat hummocks, oc- bility of glaciofl uvial, aeolian and fl uvial de- curred commonly in the areas of ground water posits can vary considerably depending on the seepage. fi ne-grain content. Moreover, some amount Air temperature is a crucial environmental of the unexplained variation in the landform, driver for frost processes although usually only particularly cryoturbation-based feature data, indirect information, such as a model of solar could be caused by the lack of information radiation balance, is available for spatial anal- on silty soils (e.g. Schunke & Zoltai 1988; van yses (e.g. Heggem et al. 2001). The radiation Vliet-Lanoë 1988a, 1991). variable utilised in this study was a simplifi ed The vegetation data utilised in this study surrogate of solar radiation but it was still a contained certain problems. The mountain fairly good approximation of the relative ra- birches under two metres high were included diation differences in the region. Moreover, an in the shrub class (Sihvo 2002: 43) and shrub empirical model of air temperature was avail- cover was determined only when the canopy able but it was not included in the fi nal set of cover was less than 30% (Eeronheimo 1996: predictors. The minimum temperature variable 11). Based on the fi eld observations, the possi- was highly correlated (Rs = 0.927) with mean ble sources of errors were, however, quite mi- altitude and could be seen as a more important nor because fi rstly the shrub cover decreased determinant for frost processes than altitude. rapidly with increasing tree cover and, second- However, the minimum temperature model ly, small mountain birches had a very scattered was based on, among others, mean altitude, distribution. and it was constructed with 34 measurements In addition to the data related shortcom- from which only fi ve were in the present study ings, collinearity between predictor variables area. Therefore, mean altitude was seen to be a (i.e. multicollinearity) can cause harm in detect- more reliable explanatory variable and thus, it ing environmental determinants (see MacNal- was chosen for the set of predictors. ly 1996). For example, insignifi cant or indirect The soil data was compiled from three variable A can omit meaningful direct factor different information sources because a fi ne- B and C from the model due to intercorrela- scale soil map was not available from the study tion (e.g. MacNally 2002: 1398). Consequently, area. The coarse-scaled Quaternary deposit if we perform spatial statistical analysis, we map (Kujansuu 1981) was completed with in- should take into account the intercorrelation formation obtained from the biotope database of the predictors to make reliable inferences and fi eld observations. In general, the utilised (e.g. Heikkinen et al. 2004). In this study, the soil data was estimated to be rather good but hierarchical partitioning approach was used to some possible sources of errors were detect- reveal the most likely causal variables that may ed. Firstly and most importantly, peat data may have remained undetected by the GLMs (see be insuffi cient because the peaty areas were the next chapter). determined based on mire vegetation (e.g. Eeronheimo 1996: 26). This could cause precariousness in several models where the peat 6.3.3 Statistical modelling variable was a critical factor. Nonetheless, the mire information layer of the topographical Statistical models are often static but processes maps and fi eld observations concord almost in nature are dynamic, and do not usually fulfi l perfectly with the utilised peat data. Secondly, the equilibrium assumption (Guisan & Zim- according to the fi eld observations, the rock mermann 2000: 153). However, periglacial terrain layer of biotope data included some processes are relatively sluggish, but above all,

122 Discussion most of the mature features are fairly stabile in sponse data, from ca. eight to 60%. In general, short time periods (in decades) and this enables the large amount of unexplained variation can statistical modelling of periglacial landforms. be caused by insuffi cient explanatory data (e.g. Statistical modelling of periglacial phenomena indirect factors or poor quality of the predic- has gained more attention from the beginning tors) or lack of causal factors from the fi nal of the 1990s although other computer-based model (e.g. Chapters 6.2.7 & 6.2.9). However, spatial models have been constructed earlier most of the logistic models explained over (Nelson 1986; Jorgenson & Kreig 1988). The 20% of the deviance change in the data that majority of the studies have concentrated on can be kept as a rather good result (e.g. Clark mapping and modelling permafrost distribu- & Hosking 1986: 463). Another reason for the tion or indicators of permafrost (e.g. Keller fairly high amount of unexplained variation 1992; Hoelzle & Haeberli 1995; Leverington & can be in the GLM method itself. Sometimes Duguay 1997; Etzelmüller et al. 1998; Gruber GLMs are not fl exible enough to capture the & Hoelzle 2001; Hoelzle et al. 2001; Luoto & shape of the interactions between environ- Seppälä 2002a; Luoto et al. 2004a; Lewkowicz mental correlates and responses (e.g. Guisan & Ednie 2004; Heggem et al. 2005) and only & Zimmermann 2000: 161). Even by adding few studies have focused on the other perigla- higher polynomial terms (e.g. a cubic term), cial phenomena (e.g. Etzelmüller et al. 2001; the approximation may still be inadequate. Luoto & Hjort 2004, 2005; Hjort & Luoto The solution to the detection of more com- 2005, 2006). plex responses could be the utilisation of non- Generalized linear models constitute a more parametric methods, such as classifi cation tree fl exible family of methods than traditional analysis (CTA) and multiple adaptive regres- least square regression techniques. GLMs han- sion splines (MARS) that allows a wider range dle non-linear relationships and different types of response curves to be modelled (e.g. Luoto of statistical distributions of geographical data & Hjort 2005). However, non-parametric tech- types, such as discrete, categorial, ordinal and niques have generally very little statistical theo- continuous data. Therefore, GLMs provide ry to support them and it is easy to over-fi t and a useful modelling framework for testing the over-explain features of the data. In addition, shapes of the response functions and signifi - non-parametric methods have been criticized cance of variables describing environmental because they can produce very complex model gradients (e.g. Franklin 1995). However, the outputs that are diffi cult to interpret (e.g. Ve- technique and data-related constrains should nables & Ripely 2002: 211). Thus, correspond- also be considered. For example, GLMs, ing parametric functions may capture most of which are based on standard regression the- the same variation and have a more realistic ory, assume that all predictors are measured explanation (Guisan & Zimmermann 2000: without error. However, the exact value of the 158–165; cf. Luoto & Hjort 2005). variables at each grid square was not known in Traditional regression techniques may dis- this study, which may introduce an additional tort inferences about the relative importance source of error into the results (e.g. Clark & of predictor variables. This is because these Hosking 1986: 317; Yee & Mitchell 1991). approaches do not take into account the inter- Furthermore, collinarity between predictors correlation of the predictors. The hierarchical and spatial autocorrelation can hamper the partitioning (HP) method used in this study detection of causal correlates by regression provides a separate measure of the amount of techniques because the presence of autocor- variation explained independently by two or relation may infl ate the degrees of freedom in more explanatory variables and, therefore, it the test of signifi cance (Chapter 6.3.1; Legen- helps to make better deductions of the impor- dre 1993; Diniz-Filho et al. 2003). tant environmental determinants (MacNally The predictors of the fi nal GLMs could 1996). However, there are two shortcomings explain only part of the variation in the re- in this approach. Firstly, HP does not produce

123 Discussion a model and, secondly, the importance of coarse modelling resolution caused the loss of polynomial variables cannot be assessed. Con- information. Secondly, the approach classifi ed, sequently, HP is suitable method in exploring according to the average environmental condi- causal factors but not, for example, in predict- tions, unsuitable sites as present squares even ing landform distributions. Moreover, com- if the feature cover was minor (0.01 ha). The putational power required for analysing large fi rst problem could be overcome by using a data set can be seen as a minor shortcoming fi ner resolution, although this could introduce too. Hence, due to the different advantages a new source of error (see Chapter 4.1.1; Hurl- and disadvantages of the GLM and HP, both bert 1984; Luoto & Hjort in press). The second methods were used and the results compared issue is of course not an error, but decreases in this study. the power of statistical model to detect the as- Many of the spatial modelling studies have sociation between response and explanatory utilised a grid-based approach (e.g. Kunin variables. This shortcoming could be removed 1998; Rowbotham & Dudycha 1998; Luoto by using a threshold to omit the squares with et al. 2002; Heikkinen et al. 2004; Ayalew & very low feature cover. However, this was not Yamagishi 2005). An advantage of the spatial performed in this study because some of the grid system is the possibility to convert usually real observations would then be removed. vacillating spatial variables to numeric form In summary, novel statistical techniques enabling statistical analysis and the possibil- combined with grid-based GIS data offers an ity to utilise GIS data as a source of predictor effi cient framework for analysing geomorpho- variables. The spatial grid system has also been logical phenomena. Numerical spatial model- used in periglacial distribution studies (e.g. ling can be an extremely useful addition to the Keller 1992; Leverington & Duguay 1997; Et- current range of techniques available to re- zelmüller et al. 2001; Luoto & Seppälä 2002a, searchers to map and monitor different land- Luoto et al. 2004a; Luoto & Hjort 2005). How- forms and processes. However, once again, it ever, examples of abundance studies in perigla- should be noted that the data related problems cial themes are largely lacking (for exception and method-based weaknesses discussed in see Luoto & Hjort 2004). The utilisation of this chapter may bias the modelling results and the grid-based method at the 25-ha resolution the model outcomes should be interpreted also posed disadvantages. Firstly, the rather critically.

124

7 SUMMARY

In this study, periglacial landforms were vided into model calibrating (70%) and evalu- mapped and analysed from an area of 600 ation (30%) sets. The random approach was km2 in subarctic Finland in the zone of dis- used to ensure the similarity of model calibra- continuous permafrost. More precisely, the tion and evaluation data. The statistical meth- distribution and abundance of the active per- ods were generalized linear modelling (GLM) iglacial features were modelled in a grid-based and hierarchical partitioning (HP). GLMs system utilising statistical methods to explore were used to produce distribution and abun- the important landscape scale environmental dance models and HP to reveal independently correlates. Empirical data sets and multivariate the most likely causal variables. techniques in combination with GIS data ena- The analyses were performed in statistical bled the construction of models to study the software R. In GLMs, the variables were se- relationship between specifi c landforms and lected using a statistically focused backward environmental factors. elimination approach. The possibility of a non- The periglacial landforms were mapped linear relationship between the explanatory with aerial photographs and located with a variable and response variable was examined by GPS-device in the fi eld. The fi eld-mapping including quadratic terms of the predictors into results were digitized in a vector format on the modelling. In the end, the prediction abili- ortho-rectifi ed aerial photographs. Periglacial ties of the GLMs were assessed with evaluation landforms were classifi ed to six main groups: data and fi eld observations. Moreover, the cor- (1) permafrost and thermokarst, (2) patterned respondence of the model variables to poten- ground, (3) slope phenomena, (4) frost weath- tial causal environmental factors was evaluated ering, (5) nival phenomena and (6) aeolian using literature and results of the HP analyses. phenomena. The whole study area was divided The binary logistic regression models were as- into equal-sized 25-ha squares. Using GIS- sessed quantitatively by calculating Kappa (κ) techniques, response variables indicating the and area under the curve (AUC) measures. landform presence/absence and abundance Abundance models were evaluated quantita- (i.e. cover) were produced in each modelling tively by calculating Spearman’s rank correla- square. tion coeffi cient (Rs) between the predicted and Twenty-four explanatory variables, which observed values. Furthermore, predictions of are potentially important in modelling perigla- the distribution models were evaluated visually cial phenomena, were compiled for each grid and empirically using both prediction maps square. The predictors were calculated from and new fi eld observations. GIS data sets and models. The variables were A total of 40 different periglacial landform classifi ed into fi ve groups based on the envi- types and subtypes were identifi ed from the ronmental factor that they refl ected; namely study area. The largest landform groups with topography, soil moisture, temperature and so- over ten types were patterned ground and lar radiation, soil type and vegetation. In addi- slope phenomena. The most common features tion, a group of spatial variables was compiled were earth hummocks, peat pounus, sorted to use in hierarchical partitioning for control- nets, sorted solifl uction sheets and defl ations. ling the effect of spatial autocorrelation and From the mapped periglacial landform types geographical trend in the response data. The twelve were utilised in distribution model- fi nal set of environmental predictors for the ling, namely palsas, convex non-sorted circles, statistical analyses was selected based on the stony earth circles, earth hummocks, peat pou- correlation analysis (Spearman’s rank correla- nus, stone pits, sorted nets, stone stripes, non- tion coeffi cient < 0.75). sorted solifl uction terraces, sorted solifl uction The fi nal modelling data were randomly di- sheets, sorted solifl uction streams and defl a-

125 Summary tions. Palsas, convex non-sorted circles, earth earth circles occurred in areas where other per- hummocks, peat pounus and sorted solifl uc- iglacial landforms were rather rare. Earth hum- tion sheets were used in abundance modelling, mocks and peat pounus had a fairly similar dis- respectively. tribution at the landscape level, although earth The potentially important environmental hummocks were more abundant and obtained factors affecting periglacial landform distribu- wider range through topographical gradients. tion and abundance are summarised in Figures On a local scale, soil properties and vegetation 115 and 116. In general, pure topographical characteristics of earth hummock and pounu factors often were the primary correlates in fi elds differed slightly. Flatness, shrub cover distribution and abundance modelling because and soil properties specifi ed the distribution they are summary surrogates for several cru- of stone pits. In proportion, topographical cial environmental determinants of periglacial factors and soil material were important cor- processes, such as temperature and soil mois- relates for sorted net occurrence. Active stone ture distribution. In addition, soil property and pits and sorted nets occupied rather the same vegetation predictors were commonly in the environments although stone pits were, in models. However, there emerged differences general, more common than nets. Slope angle between the set of important predictors for and altitude were crucial factors for the pres- the distribution and abundance modelling of ence of active sorted stripes, sorted solifl uc- the periglacial landforms. For example, canopy tion sheets and sorted solifl uction streams. cover was rather crucial discriminator between The possible importance of the vegetation occupied and unoccupied sites but it was not variables was common among the sorted slope a meaningful factor in the abundances model- features as well. Non-sorted solifl uction ter- ling. races were more abundant at lower altitudes Continuums of periglacial landforms were than sorted solifl uction features where the lo- prevalent in the study area. At lower altitudes cal topography varied considerably. Defl ations with gentle slope angles occurred earth hum- occurred commonly in sparsely vegetated sand mock, stone pit, peat pounu and palsa sequenc- and gravel areas but also on the glacigenic as es and at higher altitudes with steeper slopes well as peat landforms. Therefore, at least veg- sorted stripe, stream and sheet continuums etation and soil type information should be were common. However, sorted solifl uction used in the distribution modelling of erosional features can be without specifi c order on the aeolian features. same fell slopes. In general, palsas were more Most of the GLMs obtained good or excel- abundant below the treelimit, but they occurred lent prediction ability in the model evaluation also in regio alpina. More precisely, peat cover although the explanation power as well as the and thickness, fl at topography, soil moisture number and validity of the predictors varied and vegetation were important factors deter- notably between the fi nal models. The best dis- mining the palsa distribution and abundance tribution models were constructed with palsa, on the landscape scale. Convex non-sorted cir- earth hummock, peat pounu, sorted solifl uc- cles often located at higher altitudes than the tion sheet and sorted solifl uction stream data. other non-sorted circles, although convex cir- Stony earth circle, non-sorted solifl uction ter- cles could be spatially connected to the earth race and sorted net models had defi ciencies in hummocks and sorted nets. Important factors the prediction ability. In abundance modelling, in predicting convex non-sorted circles occur- earth hummock and peat pounu models were rence were topographical and soil moisture demonstrated to be the best (Rs > 0.7) and variables. palsa model the worst. The set of important factors in determin- The employment of a spatial grid system ing the stony earth circle presence was more with generalized linear modelling and hierar- versatile including soil property, vegetation chical partitioning proved to be a useful way and topographical predictors. Generally stony to estimate the role of potentially important

126 Summary environmental predictors in determining the predictors (non-causality), scale differences occurrence of periglacial landforms. However, between the processes and utilised predictors the utilised mapping and modelling approach (mesoscale resolution), difﬁ culties of compil- included several data and method-based short- ing causal correlates as well as analytical re- comings, which may bias the model outcomes. quirements and limitations are examples of The positional errors, interpretation difﬁ cul- the critical methodological issues that should ties and uncertainties, spatial autocorrelation, be taken into account in the different stages multicollinearity, the indirect nature of some of modelling.

Figure 115. Potentially important environmental factors affecting periglacial feature distribution in northernmost Finnish Lapland at a landscape scale. The factors were grouped into three classes based on the univariate analyses, ﬁ nal generalized linear models and HP analyses. In determining the relative importance, validity of the predictor was also taken into account (large black circle = predictor was important in GLM and HP, small black circle = predictor was important/rather important in GLM and/or HP, open small circle = predictor had some importance in GLM and/or HP). The direction of the effect is not presented because some of the factors can be positive or negative or non-linear correlates of the feature occurrence (e.g. Chapter 6.2.1).

127 Summary

Figure 116. Potentially important environmental factors affecting periglacial feature abundance in northernmost Finnish Lapland at a landscape scale (for more details see Fig. 115).

128

8 CONCLUSIONS

Periglacial landforms were common geomor- dition, geographical phenomena are typically phological features in the study area. The di- complex systems, thus a multivariate approach versity and activity of the phenomena varied is needed to explore the correlation structure spatially. The most active areas occurred com- of explanatory and response variables. monly above the treelimit in the moist valleys Thirdly, despite that many of the variables and on the fell slopes. Morphologically, per- in the models were axiomatic, the shapes of iglacial landforms resembled those described the response curves were often not. By in- from other subarctic regions. The present en- troducing second-order polynomial variables, vironmental conditions of the study area have it was possible to detect non-linear response promoted the formation of different cryo- functions of environmental factors. For exam- turbation and peat accumulation based non- ple, the several quadratic response functions sorted features. Most of the sorted landforms displayed that the suitable conditions for per- were probably formed before the climatic op- iglacial landforms are often not at the end of timum over 8000 years ago, although many of the environmental gradients, but rather at the them are active under current climate. intermediate values. To take even one step fur- The modelling approach presented in this ther, higher order polynomials could be used study provided a versatile method of model- to detect more complex responses. In addi- ling the relationship between environmental tion, thresholds and optimums of the factors correlates and periglacial landforms as well could be determined to study the potential ef- as predicting distributions and abundances fects of climate change. However, in that case, of the features. Moreover, the results of the the quality of the data should be higher and analyses concord well with the hypotheses of predictors direct factors of periglacial phe- feature occurrence in subarctic Fennoscandia nomena. at the landscape scale. Conversely, it can be ar- Fourthly, utilising simultaneously two or gued that the results of GLMs and HP only more approaches we can gain deeper insight showed what is already known. However, the into the process-environment relationships. results of this study have to be seen in a larger For example, several cryogenic landforms context. Firstly, the predictive mapping of probably have different genesis in different geomorphological landforms (and other geo- periglacial (e.g. seasonally frozen vs perma- graphical phenomenon too) in remote regions frost) regions and many of the processes are is important because distribution maps of a still poorly understood (e.g. Grab 2005: 145– different kind are essential tools to landscape 149). By modelling the phenomena through planners, conservation managers, engineers, environmental gradients, we can obtain new scientists and teachers. For example, the spa- information of environmental drivers affect- tial information of the cryogenic phenomena ing landforms genesis. is needed to monitor the possible effects of Fifthly, despite that the periglacial land- environmental change on the human activity forms were spatially autocorrelated and pre- in polar regions. dictors commonly intercorrelated, most of Secondly, numerical approaches and GIS the predictors in the fi nal models were realistic data can be used to synthesize interactions be- when compared to literature and observations tween different phenomena and environmen- in the fi eld. Therefore, the potential problems tal drivers in physical geography. For example, and constrains of geographical data sets did the occurrence of geomorphological land- not, at least in most cases, hamper the detec- forms is commonly a result of several factors tion of key environmental factors (cf. Diniz- and the relative importance of the correlates Filho et al. 2003). This gave confi dence about can be diffi cult to determine a priori. In ad- the possibility to utilise landform and GIS data

129 Conclusions in statistical analyses to identify important fac- used to obtain better models for prediction, tors of geomorphological phenomena. exploration and extrapolation purposes. A sig- Finally, too often the evaluation of the sta- nifi cant task is the compilation of valid and tistical models is neglected despite that it is a reliable explanatory variables, especially from crucial step in the numerical analyses. Models the digital elevation models and remote sens- cannot be applied confi dently without knowl- ing data and their combinations, applied to edge of their accuracy as well as the nature and model soil moisture, ground temperatures and source of the errors because the data related snow distribution in periglacial environments. problems and method-based shortcomings In statistical modelling, we should study land- may bias the modelling results. Hence, the as- form distributions on different scales because sessment of the models, especially distribution parameters and processes important at one models, was emphasised in this study. The uti- scale are frequently not important or predic- lisation of two different evaluation measures tive at another scale, and information is often and fi eld checking of the prediction maps gave lost as data are converted to coarser scales of not only a comprehensive view of the model resolution. Furthermore, the analyses should performance but also valuable information of be performed in different environments with the prediction errors. Thus, model assessment different statistical approaches taking into ac- techniques presented in this study can play an count the spatiality and multivariate nature of important role in future and it is highly recom- the geographical variables. mended that spatial statistical models should In the end, the mapping approach pre- be evaluated with different quantitative and sented in this study does not reduce the im- qualitative approaches. However, the utilisa- portance of the traditional research meth- tion of independent evaluation data would im- ods, for example the geomorphological fi eld prove the reliability of the model predictions, survey. On the contrary, it expands the pos- which is important if the models are extrapo- sibility to study complex periglacial phenom- lated to other regions. ena more effi ciently. Thorn (1992: 3), Barsch New information of, for example, occur- (1993: 158–160) and Seppälä (1997a: 83) have rence of periglacial landforms in subarctic proposed guidelines for periglacial research Finland, environmental factors, shapes of and several of the presented study problems the responses and applicability of the GLMs could be analysed further with modern spatial in predicting geomorphological phenomena mapping and modelling techniques. Cost-effi - was gained in this study. However, there are cient and reliable surveying of extensive po- still numerous unsolved problems and open lar regions, determination of the interactions questions that are related to data, scale and between processes and landforms as well as methodological issues. In future, the analy- between phenomena and environmental con- ses should focus on periglacial process units ditions, and the estimation of future modifi ca- to deepen the theoretical understanding of tion of the periglacial domain in the context the affective factors. In addition, more physi- of projected climate change are examples of cally causal environmental variables should be forthcoming challenges in periglacial studies.

130

REFERENCES

Aartolahti, T., 1969. On patterned ground in south- 23. National Board of Survey and Geographical ern Finland. Annales Academia Scientiarum Fenni- Society of Finland, Helsinki. cae, Series A3, 104. 30 pp. Atlas of Finland, 1984. Folio 112, Mapping of Fin- Aartolahti, T., 1980. The research of periglacial land. National Board of Survey & Geographical morphology in Finland (in Finnish with English Society of Finland, Helsinki. 40 pp. summary). Terra, 92: 74-87. Atlas of Finland, 1987. Folio 131, Climate. National Åhman, R., 1977. Palsas in Northern Norway (in Board of Survey & Geographical Society of Swedish with English summary). Meddelanden Finland, Helsinki. 31 pp. från Lunds Universitets Geografiska Institution, Atkinson, J., Jiskoot, H., Massari, R. & Murray, T., Avhandlingar, 78. 165 pp. 1998. Generalized linear modelling in geomor- Ahti, T., Hämet-Ahti, L. & Jalas, J., 1968. Vegeta- phology. Earth Surface Processes and Landforms, 23: tion zones and their sections in northwestern 1185–1195. Europe. Annales Botanici Fennici, 5: 169–211. Auer, V., 1924. Enontekiön suurmättäistä (palsois- Åkerman, J., 1980. Studies on periglacial geomor- ta). Terra, 36: 149–154. phology in West Spitsbergen. Meddelanden från Augustin, N., Mugglestone, M.A. & Buckland, S.T., Lunds Universitets Geografiska Institution, Avhand- 1996. An autologistic model for spatial distribu- lingar, 89. 297 pp. tion of wildlife. Journal of Applied Ecology, 33: Åkerman, J., 1993. Solifluction and creep rates 339–347. 1972–1991, Kapp Linné, West Spitsbergen. In: Ayalew, L. & Yamagishi, H., 2005. The application Frenzel, B., Matthews, J.A. & Glaeser, B. (eds.). of GIS-based logistic regression for landslide Solifluction and climate variation in the Holocene. Paleo- susceptibility mapping in the Kakuda-Yahiko klimaforschung, 11: 225–249. Mountains, Central Japan. Geomorphology, 65: Åkerman, J. & Malmström, B., 1986. Permafrost 15–31. mounds in the Abisko area, northern Sweden. Ballantyne, C.K., 1984. Nonsorted patterned ground Geografiska Annaler, 68A: 155–165. on mountains in the Northern Highlands of Albjär, G., Rehn, J. & Strömquist, L., 1979. Notes Scotland. Biuletyn Peryglacjalny, 30: 15–34. on talus formation in different climates. Geogra- Ballantyne, C.K., 2001. Measurement and theory fiska Annaler, 61A: 179–185. of ploughing boulder movement. Permafrost and Allen, P.A., 1997. Earth surface processes. Blackwell Periglacial Processes, 12: 267–288. Science, Cambridge. 404 pp. Ballantyne, C.K. & Matthews, J.A., 1982. The devel- Andersson, J.G., 1906. Solifluction, a component opment of sorted circles on recently-deglaciated of subaerial denudation. Journal of Geology, 14: terrain, Jotunheimen, Norway. Arctic and Alpine 91–112. Research, 14: 341–354. Anisimov, O. & Fitzharris, B., 2001. Polar regions Ballantyne, C.K. & Matthews, J.A., 1983. Desicca- (Arctic and Antarctic). In: McCarthy, J.J., Can- tion cracking and sorted polygon development, ziani, O.F., Leary, A.N., Dokken, D.J. & White, Jotunheimen, Norway. Arctic and Alpine Research, K.S. (eds.). Climate change 2001: Impacts, adapta- 15: 339–349. tion, and vulnerability. Contribution of working group Ballantyne, C.K. & Harris, C., 1994. The periglacia- II to the third assessment report of the Intergovernmen- tion of Great Britain. Cambridge University Press, tal Panel on Climate Change, 801–841. Cambridge Cambridge. 330 pp. University Press, Cambridge. Barsch, D., 1993. Periglacial geomorphology in the Anonymous, 2002. Digital biotope data. Metsähallitus, 21st century. Geomorphology, 7: 141–163. Helsinki. Benedict, J.B., 1970. Downslope soil movement in Arvola, E., 1987. Suomen ilmastoon vaikuttavat il- a Colorado alpine region; rates, processes, and mavirtausjärjestelmät ja säätyypit. In: Alalammi, climatic significance. Arctic and Alpine Research, P. (ed.). Atlas of Finland, Folio 131, Climate, 22– 2: 165–226.

131 References

Benedict, J.B., 1976. Frost creep and gelifluction Burrough, P.A. & McDonnell, R.A., 2000. Principles features; a review. Quaternary Research, 6: 55–76. of geographical information systems. 3rd ed. Oxford Berthling, I., Eiken, T. & Sollid, J.L., 2001. Frost University Press, New York. 306 pp. heave and thaw consolidation of ploughing Bustamante, J. & Seoane, J. 2004. Predicting the dis- boulders in a mid-alpine environment, Finse, tribution of four species of raptors (Aves: Ac- Southern Norway. Permafrost and Periglacial Proc- cipitridae) in southern Spain: statistical models esses, 12: 165–177. work better than existing maps. Journal of Bioge- Beven, K.J. & Kirkby M.J., 1979. A physically based, ography, 31: 295–306. variable contributing area model of basin hy- Carter, L.D., 1981. A Pleistocene sand sea on the drology. Hydrological Science Bulletin, 24: 43–69. Alaskan Arctic coastal plain. Science, 211: 381– Bledsoe, B.P. & Watson, C.C., 2001. Logistic analy- 383. sis of channel pattern thresholds; meandering, Chevan, A. & Sutherland, M., 1991. Hierarchical braiding, and incising. Geomorphology, 38: 281– partitioning. American Statistician, 45: 90–96. 300. Christensen, T.R., Johansson, T., Åkerman, H.J., Bocco, G., Mendoza, M. & Velázquez, A., 2001. Mastepanov, M., Malmer, N., Friborg, T., Crill, Remote sensing and GIS-based regional geo- P., & Svensson, B.H., 2004. Thawing sub-arctic morphological mapping – a tool for land use permafrost; effects on vegetation and methane planning in developing countries. Geomorphology, emissions. Geophysical Research Letters, 31. L04501, 39: 211–219. doi: 10.1029/2003GL018680. Boelhouwers, J. & Sumner, P., 2003. The palaeoen- Christiansen, H.H., 1998. Nivation forms and proc- vironmental significance of southern African esses in unconsolidated sediments, NE Green- blockfields and blockstreams. In: Phillips, M., land. Earth Surface Processes and Landforms, 23: Springman, S.M. and Arenson, L.U. (eds.). Per- 751–760. mafrost, Eight International Conferencet, Proceedings, Clapperton, C.M., 1975. Further observations on 73–78. A.A. Balkema Publishers, Lisse. the stone runs of the Falkland Islands. Biuletyn Boelhouwers, J., Holness, S., Meiklejohn, I. & Sum- Peryglacjalny, 24: 211–217. ner, P., 2002. Observations on a blockstream in Clark, M.J., 1988 (ed.). Advances in periglacial geomor- the vicinity of Sani Pass, Lesotho Highlands, phology. John Wiley & Sons, Chichester. 481 pp. Southern Africa. Permafrost and Periglacial Proc- Clark, M.J. & Seppälä, M., 1988. Slushflows in a esses, 13: 251–257. subarctic environment, Kilpisjärvi, Finnish Lap- Box, G.E.P. & Cox, D.R., 1964. An analysis of land. Arctic and Alpine Research, 20: 97–105. transformation. Journal of the Royal Statistical Soci- Clark, M.J., Gurnell, A.M., Milton, E.J., Seppälä, M. ety, 26B: 211–252. & Kyöstilä, M., 1985. Remotely-sensed vegeta- Brito, J.C., Crespo, E.G. & Paulo, O.S., 1999. Mod- tion classification as a snow depth indicator for elling wildlife distributions: logistic multiple hydrological analysis in sub-arctic Finland. Fen- regression vs overlap analysis. Ecography, 22: nia, 163: 195–216. 251–260. Clark, W.A.V. & Hosking, P.L., 1986. Statistical meth- Brown, J., Ferrians, Jr., O.J., Heginbottom, J.A. & ods for geographers. John Wiley & Sons, New York. Melnikov, E.S., 1997. Circum-arctic Map of Per- 518 pp. mafrost and Ground-Ice Conditions (scale 1:10 000 Climatological Statistics in Finland 1961–1990, 000). U.S. Geological Survey, Reston. 1991. Supplement of meteorological yearbook of Fin- Brown, R.J.E., 1970. Occurrence of permafrost in land. Finnish Meteorological Institute, Helsinki. Canadian peatlands. Proceedings of the Third Inter- 125 pp. national Peat Congress, 174–181. Department of Cohen, J., 1960. A coefficient of agreement for Energy, Mines and Resources & National Re- nominal scales. Educational and Psychological Meas- search Counsil of Canada, Ottawa. urement, 20: 37–46. Brunsden, D., 1979. Mass movements. In: Emble- Collett, D., 2003. Modelling binary data. 2nd ed. Chap- ton, C. & Thornes, J. (eds.). Process in geomorphol- man & Hall/CRC, Boca Raton. 387 pp. ogy, 130–186. Edward Arnold, London. Congalton, R.G., 1991. A review of assessing the

132 References

accuracy of classifications of remotely sensed Dobson, A.J., 2002. An introduction to generalized lin- data. Remote Sensing Environment, 37: 35–46. ear models. 2nd ed. Chapman & Hall/CRC, Baco Cook, F.A. & Raiche, V.G., 1962. Simple transverse Raton. 225 pp. nivation hollows at Resolute, NWT. Geographical Dongxing, G., Yizing, H. & Xiufeng, Z., 1993. A Bulletin, 18: 79–85. preliminary research of solifluction terraces in Cook-Talbot, J.D., 1991. Sorted circles, relative-age Fenghuoshan Pass Basin on Qinghai-Xizang dating and paleoenvironmental reconstruction Plateau. In: Guodong, C. (ed.). Permafrost, Sixth in an alpine periglacial environment, eastern International Conference, Proceedings, 210–213. Chi- Jotunheim, Norway: lichenometric and weather- nese Academy of Sciences, Lanzhou Institute ing-based approaches. The Holocene, 1: 128–141. of Glaciology and Geocryology, Lanzhou. Corte, A.E., 1966. Particle sorting by repeated Dredge, L.A., 2000. Age and origin of upland block freezing and thawing. Biuletyn Peryglacjalny, 15: fields on Melville Peninsula, eastern Canadian 176–240. Arctic. Geografiska Annaler, 82A: 443–454. Cox, D.R. & Snell, E.J., 1989. Analysis of binary data. Eeronheimo, H. (ed.), 1996. Metsähallituksen biotoop- 2nd ed. Chapman & Hall/CRC, London. 209 pikuviointiohje. Kokeiluversio 16.5.1996. Metsähalli- pp. tus, Ylä-Lapin luonnonhoitoalue, Ivalo. 47 p. Coxon, P., 1988. Remnant periglacial features on Ekholm, M., 1993. Drainage basins in Finland (in the summit of Truskmore, counties Sligo and Finnish with English summary). National Board Leitrim, Ireland. Zeitschrift für Geomorphologie, Sup- of Water and the Environment, Series A, 126. 166 plementband, 71: 81–91. pp. Crawley, M.J., 1993. GLIM for ecologists. Blackwell Elder, K. & Kattelmann, R., 1993. A low-angle Scientific Publications, Oxford. 379 pp. slushflow in the Kirgiz Range, Kirgizstan. Per- Cressie, N.A.C., 1993. Statistics for spatial data. John mafrost and Periglacial Processes, 4: 301–310. Wiley & Sons, New York. 900 pp. Embleton, C. & King, C.A.M., 1975. Periglacial geo- Curry, A.M. & Morris, C.J., 2004. Lateglacial and morphology. 2nd ed. Edward Arnold, London. 203 Holocene talus slope development and rockwall pp. retreat on Mynydd Du, UK. Geomorphology, 58: Eronen, M., 1979. The retreat of pine forest in 85–106. Finnish Lapland since the Holocene climatic Dabski, M. & Gryglewicz, E., 1998. Selected forms optimum: a general discussion with radiocar- of frost sorting in the marginal zone of Flaa- bon evidence from subfossil pines. Fennia, 157: jokull (Iceland). Biuletyn Peryglacjalny, 37: 19–34. 93–114. Dahl, R., 1966. Block fields, weathering pits and tor- ESRI, 1991. ARC/INFO user’s guide. Cell-based mod- like forms in the Narvik mountains, Nordland, eling with GRID. Analysis, display and management. Norway. Geografiska Annaler, 48A: 55–85. Environment Systems Research Institute, Inc. Dai, F.C. & Lee, C.F., 2001. Landslide characteristics Redlands, California. and slope instability modeling using GIS, Lantau Etzelmüller, B., 2000. On quantification of surface Island, Hong Kong. Geomorphology, 42: 213–228. changes using grid-based digital elevation mod- Dalgaard, P., 2002. Introductory statistics with R. els (DEM). Transactions in GIS, 4: 129–143. Springer–Verlag, New York. 267 pp. Etzelmüller, B., Berthling, I. & Sollid, J.L., 1998. Davis, N., 2001. Permafrost. A guide to frozen ground in The distribution of permafrost in Southern transition. University of Alaska Press, Fairbanks. Norway; a GIS approach. In: Lewkowicz, A.G. 351 pp. and Allard, M. (ed.). Permafrost, Seventh Interna- Demek, J., 1964. Castle koppies and tors in the tional Conference, Proceedings, Collection Nordicana Bohemian Highlands (Czechoslovakia). Biuletyn 57, 251–257. Centre d’Études Nordiques, Uni- Peryglacjalny, 14: 195–216. versité Laval, Québec. Diniz-Filho, J.A.F., Bini, L.M. & Hawkins, B.A., Etzelmüller, B., Ødegård, R.S., Berthling, I. & Sol- 2003. Spatial autocorrelation and red herrings in lid, J.L., 2001. Terrain parameters and remote geographical ecology. Global Ecology & Biogeogra- sensing data in the analysis of permafrost distri- phy, 12: 53–64. bution and periglacial processes: examples from

133 References

southern Norway. Permafrost and Periglacial Proc- Periglacial Processes, 14: 29–60. esses, 12: 79–92. Fronzek, S., Luoto, M. & Carte, T.R., 2005. Impacts Evans, I.S., 1972. General geomorphometry, de- of climate change on the distribution of palsas rivatives of altitude, and descriptive statistics. In: in the discontinuous permafrost zone of North- Chorley, R.J. (ed.). Spatial Analysis in Geomorphol- ern Europe. In: Kumke, T. (ed.). Terra Nostra. 2nd ogy, 17–90. Methuen & Co., London. European Conference on Permafrost, June 12 – June Evans, I.S., 2004. Geomorphometry. In: Goudie, 16, 2005, Potsdam, Germany, 124. A.S. (ed.). Encyclopedia of geomorphology, 435–439. Gallant, J.C. & Wilson, J.P., 2000. Primary topo- Routledge, London & New York. graphic attributes. In: Wilson, J.P. & Gallant, J.C. Fielding, A. & Bell, J., 1997. A review of methods (eds.). Terrain Analysis. Principles and Applications, for the asessment of prediction errors in con- 51–85. John Wiley & Sons, New York. servation presence/absence models. Environmen- Geiger, R., 1965. The climate near the ground. Harvard tal Conservation, 24: 38–49 University Press, Cambridge. 611 pp. Florinsky, I.V., 1998. Combined analysis of digital Geological map 1:1 000 000, 1987. Pre-Quaternary rocks. terrain models and remotely sensed data in land- Northern Fennoscandia. National Board of Survey, scape investigations. Progress in Physical Geography, Helsinki. 22: 33–60. Goldthwait, R.P., 1976. Frost sorted patterned Fogelberg, P. & Seppälä, M., 1986. Geomorfologia. ground; a review. Quaternary Research, 6: 27–35. In: Alalammi, P. (ed.). Atlas of Finland, Folio 122, Goodchild, M.F., 1986. Spatial autocorrelation. Con- Geomophology, 1–19. National Board of Survey & cepts and Techniques in Modern Geography 47. Geographical Society of Finland, Helsinki. Geo Books, Norwich. 49 pp. Franklin, J., 1995. Predictive vegetation mapping: Goodrich, L.E., 1982. The influence of snow cover geographic modelling of biospatial patterns in on the ground thermal regime. Canadian Geotech- relation to environmental gradients. Progress in nical Journal, 19: 421–432. Physical Geography, 19: 474–499. Goudie, A.S., 1995. The changing Earth. Rates of geo- Franklin, J., 1998. Predicting the distribution of morphological processes. Blackwell, Oxford. 302 pp. shrub species in southern California from cli- Grab, S., 1997. Thermal regime for a thufa apex and mate and terrain-derived variables. Journal of its adjoining depression, Mashai Valley, Lesotho. Vegetation Science, 9: 733–748. Permafrost and Periglacial Processes, 8: 437–445. Francou, B. & Bertran, P., 1997. A multivariate anal- Grab, S., 2005. Aspects of the geomorphology, ysis of clast displacement rates on stone-banked genesis and environmental significance of earth sheets, Cordillera Real, Bolivia. Permafrost and hummocks (thúfur, pounus): miniature cryo- Periglacial Processes, 8: 371–382. genic mounds. Progress in Physical Geography, 29: Francou, B., Le Mehaute, N. & Jomelli, V., 2001. 139–155. Factors controlling spacing distances of sorted Grosse, G., Schirrmeister, L., Kunitzky, V.V. & stripes in a low-latitude, alpine environment Hubberten, H.-W., 2005. The use of CORONA (Cordillera Real, 16 degrees S, Bolivia). Perma- images in remote sensing of periglacial geo- frost and Periglacial Processes, 12: 367–377. morphology: an illustration from the NE Sibe- French, H.M., 1974. Mass-wasting at Sachs Har- rian Coast. Permafrost and Periglacial Processes, 16: bour, Banks Island, N.W.T., Canada. Arctic and 163–172. Alpine Research, 6: 71–78. Gruber, S. & Hoelzle, M., 2001. Statistical modelling French, H.M., 1988. Active layer processes. In: of mountain permafrost distribution: local cali- Clark, M.J. (ed.). Advances in periglacial geomorphol- bration and incorporation of remotely sensed ogy, 151–177. John Wiley & Sons, Chichester. data. Permafrost and Periglacial Processes, 12: 69–77. French, H.M., 1996. The periglacial Environment. 2nd Gude, M., Daut, G., Dietrich, S., Maeusbacher, R., ed. Addison Wesley Longman Limited, Singa- Jonasson, C., Bartsch, A. & Scherer, D., 2002. pore. 341 pp. Towards an integration of process measure- French, H., 2003. The Development of periglacial ments, archive analysis and modelling in geo- geomorphology: 1– up to 1965. Permafrost and morphology; the Kärkevagge experimental site,

134 References

Abisko area, northern Sweden. Geografiska An- & Sons, Chichester. naler, 84A: 205-212. Harris, C., 1996. Physical modelling of periglacial Guisan, A. & Hofer, U., 2003. Predicting reptile solifluction; review and future strategy. Perma- distributions at the mesoscale: relation to cli- frost and Periglacial Processes, 7: 349–360. mate and topography. Journal of Biogeography, 30: Harris, C., Davies, M.C.R. & Coutard, J.-P., 1995. 1233–1243. Laboratory sedimentation of periglacial solifluc- Guisan, A. & Zimmermann, N.E., 2000. Predictive tion; significance of porewater pressures, mois- habitat distribution models in ecology. Ecological ture contents and undrained shear strengths Modelling, 135: 147–186. during soil thawing. Permafrost and Periglacial Proc- Guisan, A., Edwards Jr, T. & Hastle, T., 2002. Gen- esses, 6: 293–311. eralized linear and generalized additive models Harris, S.A., 1982a. Distribution of zonal perma- in studies of species distributions: setting the frost landforms with freezing and thawing ind- scene. Ecological Modelling, 157: 89–100. eces. Biuletyn Peryglacjalny, 29: 163–182. Gurney, S.D., 2001. Aspects of the genesis, geo- Harris, S.A., 1982b. Cold air drainage west of Fort morphology and terminology of palsas; peren- Nelson, British Columbia. Arctic, 35: 537–541. nial cryogenic mounds. Progress in Physical Geogra- Harris, S.A., 1986. The permafrost environment. Croom phy, 25: 249–260. Helm, London. 276 pp. Gurney, S.D. & Bartsch, A., 2005. Mapping the spa- Harris, S.A., 1994. Climatic zonality of periglacial tial distribution of geomorphological processes landforms in mountain areas. Arctic, 47: 184– in the Okstindan area of northern Norway, us- 192. ing Geomorphic Process Units as derived from Harris, S.A., 1998. Nonsorted circles on Plateau remote sensing and ground survey. Fennia, 183: Mountain, S.W. Alberta, Canada. In: Lewkowicz, 1–14. A.G. & Allard, M. (ed.). Permafrost, Seventh Inter- Hall, K., 1994. Some observations regarding sorted national Conference, Proceedings, Collection Nordicana stripes, Livingston Island, South Shetlands. Per- 57, 441–448. Centre d’Études Nordiques, Uni- mafrost and Periglacial Processes, 5: 119–126. versité Laval, Québec. Hall, K., Boelhouwers, J. & Driscoll, K., 2001. Some Harris, S.A., Cheng, G., Zhao, X. & Yongqin, D., morphometric measurements on ploughing 1998. Nature and dynamics of an active block blocks in the McGregor Mountains, Canadian stream, Kunlun Pass, Qinghai Province, Peo- Rockies. Permafrost and Periglacial Processes, 12: ple’s Republic of China. Geografiska Annaler, 219–225. 80A: 123–133. Hall, K., Thorn, C.E., Matsuoka, N. & Prick, A., Hättestrand, C., 1994. Boulder depressions in cen- 2002. Weathering in cold regions; some thoughts tral Sweden – remnants of a pre-Late Weichse- and perspectives. Progress in Physical Geography, 26: lian landscape? Geografiska Annaler, 76A: 153– 577–603. 160. Hallet, B., 1998. Measurement of soil motion in Hättestrand, C. & Stroeven, A.P., 2002. A relict sorted circles, western Spitsbergen. In: Lewko- landscape in the centre of Fennoscandian gla- wicz, A.G. & Allard, M. (ed.). Permafrost, Seventh ciation: Geomorphological evidence of minimal International Conference, Proceedings, Collection Nor- Quaternary glacial erosion. Geomorphology, 44: dicana 57, 415–420. Centre d’Études Nordiques, 127–143. Université Laval, Québec. Haugland, J.E., 2004. Formation of patterned Harris, C., 1982. The distribution and altitudinal ground and fine-scale soil development within zonation of periglacial landforms, Okstindan, two late Holocene glacial chronosequences: Norway. Zeitschrift für Geomorphologie, 26: 283– Jotunheimen, Norway. Geomorphology, 61: 287– 304. 301. Harris, C., 1987. Mechanisms of mass movement Heggem, E.S.F., Etzelmüller, B. & Berthling, I., in periglacial environments. In: Anderson, M.G. 2001. Topographic radiation balance mod- & Richards, K.S. (eds.). Slope stability; geotechnical els – sensitivity and application in periglacial engineering and geomorphology, 531–559. John Wiley geomorphology. Norsk Geografisk Tidsskrift, 55:

135 References

203–211. tioning (HP). In: Kumke, T. (ed.). Terra Nostra. Heggem, E.S.F., Juliussen, H. & Etzelmüller, B., 2nd European Conference on Permafrost, June 12 – June 2005. Mountain permafrost in Central-Eastern 16, 2005, Potsdam, Germany, 65. Norway. Norsk Geografisk Tidsskrift, 59: 94–108. Hjort, J. & Luoto, M., 2006. Modelling patterned Heikkinen, R.K. & Kalliola, R.J., 1989. Vegetation ground distribution in Finnish Lapland: an in- types and map of the Kevo nature reserve, tegration of topographical, ground and remote northernmost Finland. Kevo notes, 8. 39 pp. sensing information. Geografiska Annaler, 88A: Heikkinen, R.K., Luoto, M., Virkkala, R. & Rainio, 19–29. K., 2004. Effects of habitat cover, landscape Hjort, J. & Seppälä, M., 2003. Topographical prop- structure and spatial variables on the abundance erties of active and inactive patterned grounds of birds in an agricultural-forest mosaic. Journal in Finnish Lapland. In: Haeberli, W. & Bran- dová, D. (eds.). ICOP 2003, Permafrost, Extended of Applied Ecology, 41: 824–835. abstracts on current research and newly available infor- Helaakoski, A.R., 1912. Havaintoja jäätymisilmiöi- mation, 63-64. A.A. Balkema Publisher, Lisse. den geomorfologisista vaikutuksista. Meddelanden Hodgson, R. & Young, K.L., 2001. Preferential af Geografiska Föreningen i Finland, 9: 1–108. groundwater flow through a sorted net land- Helimäki, U.I., 1974. On the climate of Lapland, scape, Arctic Canada. Earth Surface Processes and especially on the hills (in Finnish with English Landforms, 26: 319–328. summary). In: Rapeli, P. (ed.). Climate of Lapland, Hoelzle, M. & Haeberli, W., 1995. Simulating the 23–27. The Research Society of Lapland, Ro- effects of mean annual air-temperature changes vaniemi. on permafrost distribution and glacier size: an Hétu, B. & Gray, J.T., 2000. Effects of environmen- example from the Upper Engadin, Swiss Alps. tal change on scree slope development through- Annals of Glaciology, 21: 399–405. out the postglacial period in the Chic-Choc Hoelzle, M., Mittaz, C., Etzelmüller, B. & Haeberli, Mountains in the northern Gaspe Peninsula, W., 2001. Surface energy fluxes and distribu- Quebec. Geomorphology, 32: 335–355. tion models relating to permafrost in European Hietaranta, J. & Liira, T., 1995. Lichenometry: da- Mountain areas: an overview of current devel- taing the weathering processes in the Kevo- opments. Permafrost and Periglacial Processes, 12: joki river valley, Northern Finland. Publicationes 53–68. Instituti Geographici Universitatis Turkuensis, 147: Högbom, A.G., 1905. Om s. k. jäslera och om vil- 111–121. lkoren för dess bildning. Geologiska Föreningen i Hinchliffe, S., Ballantyne, C.K. & Walden, J., 1998. Stockholm, Förhandlingar, 27: 19–36. The structure and sedimentology of relict talus, Holgrem, B. & Tenow, O., 1987. Local extreme and Trotternish, northern Skye, Scotland. Earth Sur- minima of winter air temperature in high-lati- face Processes and Landforms, 23: 545–560. tude mountainous terrain. In: Alexandersson, H. Hirvas, H., Lintinen, P. & Kosloff, P., 2000. An and Holmgren, B. (eds.). Climatological extremes in extensive permanent snowfield and the pos- the mountain, Physical background, geomorphological sible occurrence of permafrost in till in the and ecological consequences. UNGI Raport 65, 25–42. Ridnitšohkka area, Finnish Lapland. Bulletin of Uppsala Universitet, Department of Physical the Geological Society of Finland, 72: 47–56. Geography. Hirvas, H., Lintinen, P., Ojala, A.E.K. & Vanhala, Holness, S.D., 2001. The orientation of sorted H., 2005. Geological characteristics of Halti– stripes in the maritime subantarctic, Marion Ridnitšohkka region, Enontekiö, Finland. In: Island. Earth Surface Processes and Landforms, 26: Ojala, A.E.K. (ed.). Quaternary studies in the north- 77–89. ern and Arctic regions of Finland. Geological Survey of Holness, S.D., 2003. Sorted circles in the maritime Finland, Special Paper, 40: 7–12. Subantarctic, Marion Island. Earth Surface Proc- Hjort, J. & Luoto, M., 2005. Environmental deter- esses and Landforms, 28: 337–347. minants of earth hummock occurrence in Finn- Hugenholtz, C.H. & Lewkowicz, A.G., 2002. Mor- ish Lapland: a case study based on generalized phometry and environmental characteristics of linear modelling (GLM) and hierarchical parti- turf-banked solifluction lobes, Kluane Range,

136 References

Yukon Territory, Canada. Permafrost and Perigla- Jorgenson, M.T. & Kreig, R.A., 1988. A model cial Processes, 13: 301–313. for mapping permafrost distribution based on Hult, R., 1887. Die alpine Pflanzenformationen des landscape component maps and climate vari- nördlichsten Finnlands. Meddelanden af Societas ables. In: Senneset, K. (ed.). Permafrost, Fifth In- pro Fauna et Flora Fennica, 14: 153–228. ternational conference, Proceedings, 176–182. Tapir, Huovila, S., 1987a. On temperature inversions Trondheim. in Finnish Lapland. In: Alexandersson, H. & Kaitanen, V., 1969. A geographical study of the Holmgren, B. (eds.). Climatological extremes in the morphogenesis of Northern Lapland. Fennia, mountain, Physical background, geomorphological and 99: 1–85. ecological consequences. UNGI Raport 65, 43–52. Kaitanen, V., 1989. Relationships between ice-sheet Uppsala Universitet, Department of Physical dynamics and bedrock relief on dissected pla- Geography. teau areas in Finnish Lapland north 69º latitude. Huovila, S., 1987b. Pienilmasto. In: Alalammi, P. Geografiska Annaler, 71A: 1–15. (ed.). Atlas of Finland, Folio 131, Climate, 23–24. Kallio, P. & Lehtonen, J., 1973. Birch forest damage National Board of Survey & Geographical Soci- caused by Oporinia autumnata (Bkh.) in 1965–66 ety of Finland, Helsinki. in Utsjoki, N Finland. Reports from the Kevo Subar- Hurlbert, S.H., 1984. Pseudoreplication and the ctic Research Station, 10: 55–69. design of ecological field experiments. Ecological Kallio, P., Laine, U. & Mäkinen, Y., 1969. Vascular Monographs, 54: 187-211. flora on Inari Lapland. 1. Introduction and Lyc- Hustich, I., 1960. Plant geographical regions. In: opodiaceae – Polypodiaceae. Reports from the Kevo Sömme, A. (ed.). A geography of Norden, 54–62. Subarctic Research Station, 5: 1–108. J.W. Cappelens Forlag, Oslo. Kalliola, R., 1939. Pflanzensoziologische Untersu- Hutchinson, M.F., 1993. Development of a conti- chungen in der alpinen Stufe Finnische-Lapp- nent-wide DEM with applications to terrain and lands. Annales Botanici Societatis Zoologicae-Botanicae climate analysis. In: Goodchild, M.F. et al. (eds.). Fennicae Vanamo, 13: 1–328. Environmental Modeling with GIS, 392–399. Ox- Karczewski, A., 1975. Morphology and textural- ford University Press, New York. structural features of ground and hummocky Hutchinson, M.F., & Gallant, C.J., 2000. Digital moraine in the Paistunturit area, Finnish Lap- elevation models and representation of terrain land. Reports from the Kevo Subarctic Research Sta- shape. In: Wilson, J.P. & Gallant, J.C. (eds.). Ter- tion, 12: 34–44. rain Analysis. Principles and Applications, 29–50. Kärenlampi, L., 1972. Comparison between the mi- John Wiley & Sons, New York. croclimates of the Kevo ecosystem study sites Innes, J.L., 1983. Debris flows. Progress in Physical Ge- and the Kevo Meteorological Station. Reports ography, 7: 469–501. from the Kevo Subarctic Researche Station, 9: 50–65. Jaesche, P., Veit, H. & Huwe, B., 2003. Snow cover Karlén, W., 1976. Lacustrine sediments and treelim- and soil moisture controls on solifluction in an it variations as indicators of Holocene climatic area of seasonal frost, Eastern Alps. Permafrost fluctuations in Lapland. Geografiska Annaler, and Periglacial Processes, 14: 399–410. 58A: 1–34. Jahn, A., 1975. Problems of the periglacial zone. Polish Karte, J., 1979. Räumliche Abgrenzung und region- Scientific Publisher, Warsaw. 223 pp. ale Differenzierung des Periglaziärs (in Germany Jahn, A. & Siedlecki S., 1982. Periglacial phenomena with English summary). Bochumer Geographische on the Varanger Peninsula (Norway). Biuletyn Arbeiten, 35. 211 pp. Peryglacjalny, 29: 25–52. Käyhkö, J. & Pellikka, P., 1994. Remote sensing of Jensen, J.R., 2000. Remote sensing of the environment. the impact of reindeer grazing on vegetation in An earth recourse perspective. Prentice-Hall, New northern Fennoscandia using SPOT XS data. Jersey. 544 pp. Polar Research, 13: 115–124. Josefsson, M., 1988. Subalpine heaths as an indica- Käyhkö, J.A., Worsley, P., Pye, K. & Clarke, M.L., tor of a periglacial environment. Norsk Geogra- 1999. A revised chronology for aeolian activity fisk Tidsskrift, 42: 215–223. in subarctic Fennoscandia during the Holocene.

137 References

The Holocene, 9: 195–205. Kling, J., 1996. Sorted circles and polygons in Northern Kejonen, A., 1979. On solifluction sediments in the Sweden, distribution and processes. Publication A12. Muotkatunturit area, Northern-Lappland (in Earth Sciences Centre, Göteborg University. 28 Finnish with English summary). Turun yliopiston pp. maaperägeologian osaston julkaisuja, 40. 43 pp. Kling, J., 1998. The difference between sorted circle Kejonen A., 1994. Present and ancient solifluction and polygon morphology and their distribution in the Muotkatunturit and Ailigas areas, Finn- in two alpine areas, northern Sweden. Zeitschrift ish Lapland (in Finnish with English summa- für Geomorphologie, 42: 439–452. ry). In: Perttunen, M. (ed.). Uudet menetelmät Kojima, S., 1994. Relationships of vegetation, earth kvärtääritutkimuksessa. Symposio Mekrijärvellä hummocks, and topography in the High Arctic 20–21.4.1993. Acta Universitatis Ouluensis, Series environment of Canada. Proceedings of the NIPR A, Scientiae rerum naturalium, 251: 43–57. Symposium on Polar Biology, 7: 256–269. Kejonen, A., 1997. Permafrost and patterned Koskinen, R., 2005. Glacial hydrology and the de- grounds in Finland – periglacial or something glaciation of the Scandinavian ice sheet in the else. Bulletin of the Geological Society of Finland, 69: Paistunturit area of the Utsjoki region in North- 97–108. ern Finnish Lapland. In: Ojala, A.E.K. (ed.). Keller, F., 1992. Automated mapping of mountain Quaternary studies in the northern and Arctic regions permafrost using the program PERMAKART of Finland. Geological Survey of Finland, Special Pa- within the geographical information system per, 40: 41–51. ARC/INFO. Permafrost and Periglacial Processes, 3: Koster, A.E., 1988. Ancient and modern cold-cli- 133–138. mate aeolian sand deposition; a review. Journal Kershaw, G.P. & Gill, D., 1979. Growth and decay of Quaternary Science, 3: 69–83. of palsas and peat plateaus in the Macmillan Kotilainen, M., 2004. Dune stratigraphy as an indi- Pass-Tsichu River area, Northwest Territories, cator of Holocene climatic change and human Canada. Canadian Journal of Earth Sciences, 16: impact in northern Lapland, Finland. Annales 1362–1374. Academiae Scientiarum Fennicae, Geologica-Geograp- Kessler, M.A. & Werner, B., 2003. Self-organiza- hica, 166. 156 pp. tion of sorted patterned ground. Science, 299: Kujansuu, R., 1981. General geological map of Finland 380–383. 1:400 000. Quaternary Deposits. Sheet 39+49. Geo- Kessler, M.A., Murray, A.B., Werner, B.T. & Hallet, logical Survey of Finland, Espoo. B., 2001. A model for sorted circles as self-or- Kukkonen, I.T. & Safanda, J., 2001. Numerical ganized patterns. Journal of Geophysical Research, modelling of permafrost in bedrock in northern B, Solid Earth and Planets, 106: 13287–13306. Fennoscandia during the Holocene. Global and King, L. & Seppälä, M., 1987. Permafrost thickness Planetary Change, 29: 259–273. and distribution in Finnish Lapland – Results Kultti, S., 2004. Holocene changes in treelines and climate of Geoelectrical soundings. Polarforshung, 57: from Ural mountains to Finnish Lapland. Helsinki 127–147. University Printing House, Helsinki. 33 pp. King, L. & Seppälä, M., 1988. Permafrost sites in Kumar, L., Skidmore, A.K. & Knowles, E., 1997. Finnish Lapland and their environment. In: Sen- Modelling topographic variation in solar radia- neset, K. (ed.). Permafrost, Fifth, International con- tion in a GIS environment. International Journal ference, Proceedings, 59–61. Tapir, Trondheim. for Geographical Information Science, 11: 475–497. Klein, A., 2004. Remote sensing in geomorphology. Kunin, W.E., 1998. Extrapolating species abundance In: Goudie, A.S. (ed.). Encyclopedia of geomorphol- across spatial scales. Science, 281: 1513–1515. ogy, 844–846. Routledge, London & New York. Lagerbäck, R. & Rodhe, L., 1986. Pingos and palsas Kleman, J. & Borgström, I., 1990. The boulder in northernmost Sweden; preliminary notes on fields of Mt. Fulufjället, west-central Sweden; recent investigations. Geografiska Annaler, 68A: late Weichselian boulder blankets and intersta- 149–154. dial periglacial phenomena. Geografiska Annaler, Landis, J. & Koch, G., 1977. The measurement of 72A: 63–78. observer agreement for categorical data. Biomet-

138 References

rics, 33: 159–174. Łoziński, W. von, 1909. Über die mechanische Ver- Lappalainen, E. & Hänninen, P., 1993. The peat witterung der Sandstein im gemässigten Klima. reserves of Finland (in Finnish with English Bulletin International de l’Académie des Sciences de summary). Geological Survey of Finland, Report of Cracovie, Classe des Sciences Mathématiques et Na- Investigation, 117. 118 pp. turelles, 1: 1–25. Larsson, S., 1982. Geomorphological effects on the Lundqvist, G., 1949. The orientation of the block slopes of Longyear Valley, Spitsbergen, after a material in certain species of flow earth. Ge- heavy rainstorm in July 1972. Geografiska An- ografiska Annaler, 31A: 335–347. naler, 64A: 105–125. Lundqvist, G., 1951. Blocksänkor och några andra Lautridou, J-P., Ozouf, J.C., 1982. Experimen- frostfenomen. Geologiska Föreningen i Stockholm, tal frost shattering; 15 years of research at the Förhandlingar, 73: 505-512. Centre de geomorphologie du CNRS. Progress in Lundqvist, J., 1962. Patterned ground and related Physical Geography, 6: 215-232. frost phenomena in Sweden. Sveriges Geologiska Legendre, P., 1993. Spatial autocorrelation: trouble Undersökning, Serie C, Avhandlingar och Uppsatser, or new paradigm? Ecology, 74: 1659–1673. 583. 101 pp. Leverington, D.W. & Duguay, C.R., 1997. A neural Lundqvist, J., 1969. Earth and ice mounds; a ter- network method to determine the presence or minological discussion. In: Péwé, T.L. (ed.). The absence of permafrost near Mayo, Yukon Ter- periglacial environment, past and present, 203–215. ritory, Canada. Permafrost and Periglacial Processes, McGill–Queen`s University Press, Montreal. 8: 207–217. Luoto, M. & Heikkinen, R., 2003. Recent develop- Lewis, W.V., 1939. Snow-patch erosion in Iceland. ments in spatial methods and data in biogeo- Geographical Journal, 94: 153–161. graphical distribution modelling – advantages Lewkowicz, A.G., 1988. Slope processes. In: Clark, and pitfalls. Fennia, 181: 35–48. Luoto, M. & Hjort, J., 2004. Generalized linear M.J. (ed.). Advances in periglacial geomorphology, models in periglacial studies: terrain parameters 325–370. John Wiley & Sons, Chichester. and patterned ground. Permafrost and Periglacial Lewkowicz, A.G., & Ednie, M., 2004. Probability Processes, 15: 327–338. mapping of mountain permafrost using the BTS Luoto, M. & Hjort, J., 2005. Evaluation of current method, Wolf Creek, Yukon Territory, Canada. statistical approaches for predictive geomorpho- Permafrost and Periglacial Processes, 15, 67–80. logical mapping. Geomorphology, 67: 299–315. Lewkowicz, A.G. & Harris, C., 2005. Frequency and Luoto, M. & Hjort, J., in press. Scale matters - a magnitude of active-layer detachment failures multi-resolution study of the determinants of in discontinuous and continuous permafrost, patterned ground activity in subarctic Finland. northern Canada. Permafrost and Periglacial Proc- Geomorphology. esses, 16: 115–130. Luoto, M. & Seppälä, M., 2000. Summit peats (‘peat Lichstein, J.W., Simons, T.R., Shriner, S.A. and cakes’) on the fells of Finnish Lapland: conti- Franzreb, K.E., 2002. Spatial autocorrelation nental fragments of blanket mires? Holocene, 10: and autoregressive models in ecology. Ecological 229-241. Monographs, 72: 445-463. Luoto, M. & Seppälä, M., 2002a. Modelling the Liira, T. & Hietaranta, J., 1998. On the postglacial distribution of palsas in Finnish Lapland with development of the bedrock precipices and talus logistic regression and GIS. Permafrost and Perigla- formations in the Keävvu river valley northern cial Processes, 13: 17–28. Finnish Lapland. Reports from the Kevo Subarctic Luoto, M. & Seppälä, M., 2002b. Characteristics Research Station, 22: 1–10. of earth hummocks (pounus) with and without Linton, D.L., 1955. The problem of tors. Geographi- permafrost in Finnish Lapland. Geografiska An- cal Journal, 121: 470–487. naler, 84A: 127–136. Liston, G.E., McFadden, J.P., Sturm, M. & Pielke, Luoto M, & Seppälä, M., 2003. Thermokarst ponds R.A., Sr, 2002. Modelled changes in arctic tun- as indicators of the former distribution of pal- dra snow, energy and moisture fluxes due to in- sas in Finnish Lapland. Permafrost and Periglacial creased shrubs. Global Change Biology, 8: 17–32. Processes, 14: 19–27.

139 References

Luoto, M., Fronzek, S. & Zuidhoff, F., 2004a. Spa- block fields (in Swedish with English summary). tial modelling of palsa mires in relation to cli- Meddelanden från Lunds Universitets Geografiska In- mate in northern Europe. Earth Surface Processes stitution, Avhandlingar, 93. 351 pp. and Landforms, 29: 1373–1387. Mann, D., 2003. On patterned ground. Science, 299: Luoto, M., Heikkinen, R.K. & Carter, T.R., 2004b. 354–355. Loss of palsa mire in Europe and biological Mansikkaniemi, H., 1970. Deposits of sorted mate- consequences. Environmental Conservation, 31: rial in the Inarijoki-Tana river valley in Lapland. 30–37. Reports from the Kevo Subarctic Research Station, 6: Luoto, M., Toivonen, T. & Heikkinen, R.K., 2002. 1–63. Prediction of total and rare plant species rich- Mansikkaniemi, H., 1972. Flood deposits, transport ness in agricultural landscapes from satellite im- distances and roundness of loose material in the ages and topographic data. Landscape Ecology, 17: Tana river valley, Lapland. Reports from the Kevo 195–217. Subarctic Researche Station, 9: 15–23. Luoto, M., Kuussaari, M., Rita, H., Salminen, J. & Mansikkaniemi, H. & Laitinen, T., 1990. Pattern von Bonsdorff T., 2001. Determinants of dis- of local wind changes in a fell region, northern tribution and abundance in the clouded apollo Finland. Reports from the Kevo Subarctic Researche butterfly: a landscape ecological approach. Ecog- Station, 21: 11–20. raphy, 24: 601–617. Map of Quaternary geology 1:1 000 000, 1988. Sheet 3. Maarleveld, G., 1960. Wind direction and cover Ice flow indicators. Northern Fennoscandia. National sands in the Netherlands. Biuletyn Peryglacialny, Land Survey of Sweden, Stockholm. 8: 49–58. Mark, D.M., 1975. Geomorphometric parameters: a Mackay, J.R., 1980. The origin of hummocks, west- review and evaluation. Geografiska Annaler, 57A: ern Arctic coast, Canada. Canadian Journal of 165–177. Earth Sciences, 17: 996–1006. Matsumoto, H. & Ishikawa, M., 2002. Gelifluction Mackay, J.R. & Mackay, D.K., 1976. Cryostatic pres- within a solifluction lobe in the Kärkevagge Val- sures in nonsorted circles (mud hummocks), ley, Swedish Lapland. Geografiska Annaler, 84A: Inuvik, Northwest Territories. Canadian Journal 261–266. of Earth Sciences, 13: 889–897. Matsuoka, N., 2001a. Solifluction rates, processes MacNally, R., 1996. Hierarchical partitioning as an and landforms; a global review. Earth-Science Re- interpretative tool in multivariate inference. Aus- views, 55: 107–134. tralian Journal of Ecology, 21: 224–228. Matsuoka, N., 2001b. Microgelivation versus mac- MacNally, R., 2000. Regression and model-build- rogelivation; towards bridging the gap between ing in conservation biology, biogeography and laboratory and field frost weathering. Permafrost ecology: the distinction between – and reconcili- and Periglacial Processes, 12: 299–313. ation of – ‘predictive’ and ‘explanatory’ models. Matsuoka, N., 2004. Solifluction. In: Goudie, A.S. Biodiversity and Conservation, 9: 655-671. (ed.). Encyclopedia of geomorphology, 984–987. MacNally, R., 2002. Multiple regression and infer- Routledge, London & New York. ence in ecology and conservation biology: fur- Matsuoka, N., 2005. Temporal and spatial variation ther comments on identifying important pre- in periglacial soil movements on alpine crest dictor variables. Biodiversity and Conservation, 11: slope. Earth Surface Processes and Landforms, 30: 1397–1401. 41–58. MacNally, R. & Horrocks, G., 2002. Relative influ- Matsuoka, N., Abe, M. & Ijiri, M., 2003. Differential ence of patch, landscape and historical factors frost heave and sorted patterned ground; field on birds in an Australian fragmented landscape. measurements and a laboratory experiment. Journal of Biogeography, 29: 395–410. Geomorphology, 52: 73–85. Malmström, B. & Palmer, O., 1984. Glacial and Matthews, J.A. & Briffa, K.R., 2005. The ´Little Ice periglacial geomorphology on the Varanger Pe- Age`: re-evaluation of an evolving concept. Geo- ninsula, Northern Norway; geomorphological grafiska Annaler, 87A: 17–36. mapping with an analysis of glacial forms and Matthews, J.A., Seppälä, M. & Dresser, Q., 2005.

140 References

Holocene solifluction, climate variation and erscheinungen in Finnmark, Norwegische fire in a subarctic landscape at Pippokangas, Lappland (with English summary). Courier For- Finnish Lapland, based on radiocarbon-dated schungsinstitut Senckenberg, 143: 69–173. buried charcoal. Journal of Quaternary Science, 20: Meriläinen, K., 1965. General geological map of Finland, 533–548. Pre-Quaternary rocks, Scale 1:400 000, Sheet C 8–9, Matthews, J.A., Ballantyne, C.K., Harris, C. & Mc- Inari–Utsjoki. Geological Survey. Helsinki. Carroll, D., 1993. Solifluction and climate vari- Meriläinen, K., 1976. The granulite complex and ation in the Holocene; discussion and synthe- adjacent rock in Lapland, northern Finland. Bul- sis. In: Frenzel, B., Matthews, J.A. & Glaeser, B. letin of Geological Survey of Finland, 281. 129 pp. (eds.). Solifluction and climate variation in the Holocene. Metz, C.E., 1978. Basic principles of ROC analysis. Paleoklimaforschung, 11: 339–361. Seminars in Nuclear Medicine, 8: 720–733. Matthews, J.A., Dahl, S.-O., Berrisford, M.S. & Nesje, Mikkola, E., 1932. On the physiography and late- A., 1997. Cyclic development and thermokarstic glacial in northern Lapland. Bulletin de la Commis- degradation of palsas in the mid-alpine zone at sion géologique de Finlande, 96. 88 pp. Leipullan, Dovrefjell, southern Norway. Perma- Miettinen, S., 2002. GPS käsikirja. Genimap, frost and Periglacial Processes, 8: 107–122. Hämeenlinna. 165 pp. Matthews, J.A., Shakesby, R.A., Berrisford, M.S. & Mikkola, E. 1941. Kivilajikartan selitys. Lehdet B7 – C7 McEwen, L.J., 1998. Periglacial patterned ground – D7, Muonio – Sodankylä – Tuntsajoki. Suomen geo- on the Styggedalsbreen glacier foreland, Jotun- loginen yleiskartta, 1:400 000. Suomen geologinen heimen, southern Norway: micro-topographic, toimikunta, Helsinki. 287 pp. paraglacial and geoecological controls. Permafrost Mooers, H.D. & Glaser, P.H., 1989. Active patterned and Periglacial Processes, 9: 147–166. ground at sea level, Fourchu, Nova Scotia, Can- McBride, G.B., Loftis, J.C. & Adkins, N.C., 1993. ada. Arctic and Alpine Research, 21: 425–432. What do significance tests really tell us about Moore, I.D., Grayson, R.B. & Ladson, A.R., 1991. the environment? Environmental Management, 17: Digital terrain modelling: a review of hydrologi- 423–432. cal, geomorphological, and biological applica- McCullagh, P. & Nelder, J.A., 1989. Generalized linear tions. Hydrological Processes, 5: 3–30. models. 2nd ed. Chapman & Hall, New York. 511 Moore, P.D., 1984. Rise and fall of palsa mounds. pp. Nature, 310: 188. McGreevy, J.P., 1981. Perspectives of frost shatter- Muir, M.P., 1983. The role of pre-existing, corru- ing. Progress in Physical Geography, 5: 56–75. gated topography in the development of stone Mckay, G.A. & Gray, D.M., 1981. The distribution stripes. In: Miller, K.D. (ed.). Permafrost, Fourth of snowcover. In: Gray, D.M. & Male, D.H. International conference, Proceedings. 877–882. Na- (eds.). Handbook of Snow, 154–166. Pergamon tional Academy Press, Washington, D.C. Press, Toronto. Muller, S.W., 1947. Permafrost or permanently frozen McPherson, J., Jetz, W. & Rogers, D., 2004. The ef- ground and related engineering problems. J.W. Edwards fects of species’ range sizes on the accuracy of Inc., Michigan. 231 pp. distribution models: ecological phenomenon or National Land Survey of Finland, 2004a. OrtoCD. statistical artefact? Journal of Applied Ecology, 41: 5th of October 2004. 811–823. stability of thawing slopes. Canadian Geotechnical National Land Survey of Finland, 2004b. Height Journal, 11: 447–469. and depth information. 12th of October 2004. Meier, K.-D., 1987. Studies of periglacial landscape Germany with English summary). Jahrbuch der Nelder, J.A. & Wedderburn, R.W.M., 1972. General- Geographischen Gesellschaft zu Hannover. Sonderheft, ized linear models. Journal of the Royal Statistical 13. 298 pp. Society, 135A: 370–384. Meier, K.-D., 1991. Permafrost- und Periglaziär- Nelson, F.E., 1986. Permafrost distribution in cen-

141 References

tral Canada: application of a climate-based pre- Jotunheimen, southern Norway. In: Senneset, K. dictive model. Annals of the Association of Ameri- (ed.). Permafrost, Fifth International conference, Pro- can Geographers, 76: 550–569. ceedings, 59–61. Tapir, Trondheim. Nelson, F.E., 2003. (Un)frozen in time. Science, 299: Ohlson, B., 1964. Frostaktivität, Verwitterung und 1673–1675. Bodenbildung in den Fjeldgegenden von Enon- Nelson, F.E., Anisimov, O.E. & Shiklomanov, N.I., tekiö, Finnisch-Lappland. Fennia, 89. 180 pp. 2001. Subsidence risk from thawing permafrost. Okko, V., 1954. Periglacial morphology in Finland Nature, 410: 889–890. (in Finnish with English summary). Terra, 66: Nelson, F.E., Frederick, E., Hinkel, K.M., & Out- 54–57. calt, S.I., 1992. Palsa-scale frost mounds. In: Oksanen, J., 2003. R: Opas ekologeille. Versio 1.0. Dixon, J.C. & Abrahams, A.D. (eds.). Periglacial 22nd of October 2003. geomorphology, 305–325. Wiley, Chichester. fields in southern Norway; implications for the Oksanen, J. & Jaakkola, O., 2000. Interpolation and geometry, thickness, and isostatic loading of the accuracy of contour-based raster DEMs. Reports late Weichselian Scandinavian ice sheet. Journal of the Finnish Geodetic Institute, 2000:2. 32 pp. of Quaternary Science, 5: 225–234. Oksanen, L. & Virtanen, R., 1995. Topographical, Nesje, A., Kvamme, M. & Rye, N., 1989. Neoglacial altitudinal and regional patterns in continental gelifluction in the Jostedalsbreen region, west- and suboceanic heath vegetation of northern ern Norway; evidence from data buried palaeo- Fennoscandia. Acta Botanica Fennica, 153: 1–80. podsols. Earth Surface Processes and Landforms, 14: Onesti, L., 1985. Meteorological conditions that 259–270. initiate slushflows in the Central Brooks Range, Nicholls, A.O., 1989. How to make biological sur- Alaska. Annals of Glaciology, 6: 23–25. veys go further with generalized linear models. Østrem, G., 1971. Rock glaciers and ice-cored mo- Biological Conservation, 50: 51–75. raines, a reply to D. Barsch. Geografiska Annaler, Nicholson, F.H., 1976. Patterned ground formation 53A: 207–213. and description as suggested by Low Arctic and Overduin, P.P., Ping, C.L. & Kane, D.L., 2003. Frost Subarctic examples. Arctic and Alpine Research, 8: boils, soil ice content and apparent thermal dif- 329–342. fusivity. In: Phillips M., Springman, S.M. & Niessen, A., Horssen, P. & Koster, E.A., 1992. Al- Arenson, L.U. (eds.). Permafrost, Eight internation- titudinal zonation of selected geomorphological al conference, Proceedings, 869–874. A.A. Balkema phenomena in an alpine periglacial area (Abisko, Publishers, Lisse. northern Sweden). Geografiska Annaler, 74A: Pearce, J. & Ferrier, S., 2000. Evaluating the predic- 83–196. tive performance of habitat models using logis- Niessen, A., Koster, E.A. & Galloway, J.P., 1984. tic regression. Ecological Modelling, 133: 225–245. Periglacial sand dunes and eolian sand sheets; Peltier, L.C., 1950. The geographic cycle in perigla- an annotated bibliography. U.S. Geological Survey, cial regions as it is related to climatic geomor- Open-File Report, 84-0167. 61 pp. phology. Annals of Association of American Geog- Nyberg, R., 1985. Debris flows and slush avalanches raphers, 40: 214–236. in northern Swedish Lappland; distribution and Pereira, J.M. & Itami, R.M., 1991. GIS-based habi- geomorphological significance. Meddelanden från tat modeling using logistic multiple regression: Lunds Universitets Geografiska Institution, Avhan- a study of the Mt. Graham red squirrel. Pho- dlingar, 97. 222 pp. togrammetric Engineering and Remote Sensing, 57: Nyberg, R., 1989. Observations of slushflows and 1475–1486. their geomorphological effects in the Swedish Peterson, R.A. & Krantz, W.B., 2003. A mechanism mountain area. Geografiska Annaler, 71A: 185– for differential frost heave and its implications 198. for patterned-ground formation. Journal of Gla- Ødegård, R.S., Listøl, O. & Sollid, J.L., 1988. Per- ciology, 49: 69–80. iglacial forms related to terrain parameters in Peterson, R.A., Walker, D.A., Romanovsky, V.E.,

142 References

Knudson, J.A., Raynolds, M.K. & Krantz, W.B., Rapp, A., 1960. Recent development of moun- 2003. A differential frost heave model; cryotur- tain slopes in Kärkevagge and surroundings, bation-vegetation interactions. In: Phillips, M., northern Scandinavia. Geografiska Annaler, 42A: Springman, S.M. & Arenson, L.U. (eds.). Perma- 65–200. frost, Eight international conference, Proceedings, 21– Rapp, A., 1986. Slope processes in high latitude 25, 885–890. A.A. Balkema Publishers, Lisse. mountains. Progress in Physical Geography, 10: Péwé, T.L., 1955. Origin of the upland silt near 53–68. Fairbanks, Alaska. Geological Society of America Raup, H.M., 1965. The structure and development Bulletin, 66: 699–724. of turf hummocks in the Mesters Vig District, Piirola, J., 1969. Frost-sorted block concentrations Northeast Greenland. Meddelelser om Grønland, in western Inari, Finnish Lapland. Fennia, 99: 166. 112 pp. 5–35. Ray, R., Krantz, W., Caine, N. & Gunn, R., 1983. A Piirola, J., 1972. Patterned ground in northern Fin- model for patterned ground regularity. Journal of land (in Finnish with English summary). Terra, Glaciology, 29: 317–337. 84: 132–142. Reid, J.R. & Nesje, A., 1988. A giant ploughing Pike, R.J. & Wilson, S.E., 1971. Elevation-relief ra- block, Finse, Southern Norway. Geografiska An- tio, hypsometric integral, and geomorphic area- naler, 70A: 27–33. altitude analysis. Geological Society of America Bul- Rissing, J.M. & Thorn, C.E., 1985. Particle size and letin, 82: 1079–1084. clay mineral distributions within sorted and Ping, C.L., Michaelson, G.J., Overduin, P.P. & Stiles, nonsorted circles and the surrounding parent C.A., 2003. Morphogenesis of frost boils in the material, Niwot Ridge, Front Range, Colorado, Galbraith Lake area, Arctic Alaska. In: Phillips, U.S.A. Arctic and Alpine Research, 17: 153–163. M., Springman, S:M. & Arenson. L.U. (eds.). Rönkkö, M. & Seppälä, M., 2003. Surface character- Permafrost, Eight international conference, Proceedings, istics affecting active layer formation in palsas, 897–900. A.A. Balkema Publishers, Lisse. Finnish Lapland. In: Phillips, M., Springman, Pissart, A., 1993. Understanding the controls on S.M. & Arenson, L.U. (eds.). Permafrost, Eight In- solifluction movements in different environ- ternational Conference, Proceedings, 995–1000. A.A. ments; a methodology and its application in the Balkema Publishers, Lisse. French Alps. In: Frenzel, B., Matthews, J.A. & Rowbotham, D.N. & Dudycha, D., 1998. GIS mod- Glaeser, B. (eds.). Solifluction and climate variation in elling of slope stability in Phewa Tal watershed, the Holocene. Paleoklimaforschung, 11: 209–215. Nepal. Geomorphology, 26: 151–170 Pissart, A., 2004. Thermokarst. In: Goudie, A.S. Rudberg, S., 1977. Periglacial zonation in Scandi- (ed.). Encyclopedia of geomorphology, 1049–1051. navia. In: Poser, H. (ed.). Formen, Formengesells- Routledge, London & New York. chaften und Untergrenzen in den hötigen periglazialen Priesnitz, K., 1988. Cryoplanation. In: Clark, M.J. Hoehenstufen der Hochgebirge Europas und Afrikas (ed.). Advances in periglacial geomorphology, 49–67. zwischen Arktis und Aequator. Abhandlungen der John Wiley & Sons, Chichester. Akademie der Wissenschaften in Göttingen, Mathema- Pullan, R.A., 1959. Tors. Scottish Geographical Maga- tisch-Physikalische Klasse, 31: 92–104. zine, 75: 51–55. Rushton, S.P., Ormerod, S.J. & Kerby, G., 2004. Punkari, M., 1996. Glacial dynamics of the northern New paradigms for modelling species distribu- European ice sheets. Department of Geology, Di- tions? Journal of Applied Ecology, 41: 193–200. vision of Geology and Palaentology, University Ruuhijärvi, R., 1960. Über die regionale Einteilung of Helsinki, Helsinki. 24 pp. der nordfinnishen Moore. Annales Botanici So- Pye, K., 1984. Loess. Progress in Physical Geography, cietatis Zoologicae-Botanicae Fennicae Vanamo, 31: 8: 176–217. 1–360. Railton, J.B. & Sparling, J.H., 1973. Preliminary Ruuhijärvi, R., 1962. Palsasoista ja niiden mor- studies on the ecology of palsas mounds in fologiasta siitepölyanalyysin valossa. Terra, 74: northern Ontario. Canadian Journal of Botany, 51: 58–68. 1037–1044. Salmi, M., 1970. Development of palsas in Finn-

143 References

ish Lapland. Proceedings of the Third International 71. Reports from the Kevo Subarctic Research Station, Peat Congress, Quebec, Canada 1968, 182–189. 13: 1–11. Department of Energy, Mines and Resources & Seppälä, M., 1976b. Seasonal thawing of a palsa at National Research Counsil of Canada, Ottawa. Enontekio, Finnish Lapland, in 1974. Biuletyn Salmi, M., 1972. Present development stage of pal- Peryglacjalny, 26: 17–24. sas in Finland. Proceedings of the 4th International Seppälä, M., 1979a. Recent palsa studies in Finland. Peat Congress, Otanemi, Finland, June 25 – 30, 1972, In: Vasari, Y., Saarnisto, M. & Seppälä, M. (eds.). 121–141. Paleohydrology of the temperate zone. Acta Sass, O. & Wollny, K., 2001. Investigations regarding Universitatis Ouluensis, Series A, 82: 81–87. alpine talus slopes using ground-penetrating ra- Seppälä, M., 1979b. A new technique to measure dar (GPR) in the Bavarian Alps, Germany. Earth the rate of mass movements on slopes. Studia Surface Processes and Landforms, 26: 1071–1086. Geomorphologica Carpato-Balcanica, 13: 221–223. Schlyter, P., 1995. Ventifacts as palaeo-wind indica- Seppälä, M., 1980. Deglaciation and glacial lake de- tors in southern Scandinavia. Permafrost and Per- velopment in the Kaamasjoki river basin, Finn- iglacial Processes, 6: 207–219. ish Lapland. Boreas, 9: 311–319. Schönhage, W., 1969. Note on the ventifacts in the Seppälä, M., 1982a. An experimental study of the Netherlands. Biuletyn Peryglacialny, 20: 355–360. formation of the palsas. In: French, H.M. (ed.). Schunke, E. & Zoltai, C., 1988. Earth hummocks 4th Canadian Permafrost Conference, Proceedings, (thufur). In: Clark, M.J. (ed.). Advances in Perigla- 36–42. National Research Council of Canada, cial Geomorphology, 231–245. John Wiley & Sons, Ottawa. Chichester. Seppälä, M., 1982b. Present-day periglacial phe- Segurado, P. & Araújo, M., 2004. An evaluation of nomena in northern Finland. Biuletyn Perygla- methods for modelling species distributions. cialny 29, 231–243. Journal of Biogeography, 31: 1555–1569. Seppälä, M., 1983. Seasonal thawing of palsas in Selby, M.J., 1972. Antarctic tors. Zeitschrift für Geo- Finnish Lapland. In: Miller, K.D. (ed.). Perma- morphologie, Supplementband, 13: 73–86. frost, Fourth International conference, Proceedings, Seppä, H. & Birks, H.J.B., 2002. Holocene climate 1127–1132. National Academy Press, Washing- reconstruction from the Fennoscandian tree- ton, D.C. line area based on pollen data from Toskaljavri. Seppälä, M., 1984. Deflation measurements on Hie- Quaternary Research, 57: 191–199. tatievat, Finnish Lapland, 1974-77. In: Olson, R., Seppälä, M., 1971. Evolution of eolian relief of the Geddes, F. & Hastings, R. (eds.). Northern ecology Kaamasjoki-Kiellajoki river basin in Finnish La- and resource management, 39–49. The University of pland. Fennia, 104. 88 pp. Alberta Press, Edmonton. Seppälä, M., 1972a. The term ‘palsa’. Zeitschrift für Seppälä, M., 1986. The origin of palsa. Geografiska Geomorphologie, 16: 463. Annaler, 68A: 141–147. Seppälä, M., 1972b. Peat at the top of Ruohttir fell, Seppälä, M., 1987. Periglacial phenomena of north- Finnish Lapland. Reports from the Kevo Subarctic ern Fennoscandia. In: Boardman, J. (ed.). Per- Research Station, 9: 1–6. iglacial processes and landforms in Britain and Ireland, Seppälä, M., 1972c. Location, morphology and ori- 45–55. Cambridge University Press, Cambridge. entation of inland dunes in northern Sweden. Seppälä, M., 1988. Palsas and related forms. In: Geografiska Annaler, 54A: 85–104. Clark, M.J. (ed.). Advances in periglacial geomorphol- Seppälä, M., 1974. Some quantitative measurements ogy, 247–278. John Wiley & Sons, Chichester. of the present-day deflation on Hietatievat, Seppälä, M., 1990. Depth of snow and frost on a Finnish Lapland. Abhandlungen der Akademie der palsa mire, Finnish Lapland. Geografiska Annaler, Wissenschaften in Göttingen, Mathematisch- Physika- 72A: 191–201. lische Klasse III, 29: 208–220. Seppälä, M., 1993a. Climbing and falling sand dunes Seppälä, M., 1976a. Periglacial character of the cli- in Finnish Lapland. In: Pye, K. (ed.). The Dynam- mate of the Kevo region (Finnish Lapland) on ics and Environmental Context of Aeolian Sedimen- the basis of meteorological observations 1962– tary Systems. Geological Society Special Publication, 72:

144 References

269–274. esses. In: Seppälä, M. (ed.).The physical geography Seppälä, M., 1993b. Solifluction in northern Finland: of Fennoskandia, 59–75. Oxford University Press, past and present. In: Frenzel, B., Matthews, J.A. Oxford. & Glaeser, B. (eds.). Solifluction and climate variation Seppälä, M., 2005c. Frost heave on earth hummocks in the Holocene. Paleoklimaforschung, 11: 59–70. (pounus) in Finnish Lapland. Norsk Geografisk Seppälä, M., 1994a. The origin of palsas; an exam- Tidssktift, 59: 171–176. ple of how theories are formed (in Finnish with Seppälä, M. & Hassinen, S., 1997. Freeze-thaw in- English summary). Terra, 106: 204–208. dices in northernmost Fennoscandia according Seppälä, M., 1994b. Snow depth controls palsa to meteorological observations, 1980–1991. In: growth. Permafrost and Periglacial Processes, 5: 283– Knutsson, S. (ed.). Ground Freezing 97, Frost action 288. in soils, Proceedings of the International Symposium on Seppälä, M., 1995a. How to make a palsa: a field ex- Ground Freezing and Frost Action in Soils, 153–160. periment on formation of permafrost. Zeitschrift Balkema, Rotterdam. für Geomorphologie, Supplementband, 99: 91–96. Seppälä, M. & Rastas, J., 1980. Vegetation map of Seppälä, M., 1995b. Deflation and redeposition of northernmost Finland with special reference to sand dunes in Finnish Lapland. Quaternary Science subarctic forest limits and natural hazards. Fen- Reviews, 14: 799–809. nia, 158: 41–61. Seppälä, M., 1997a. Introduction to the periglacial Shilts, W.W., 1978. Nature and genesis of mudboils, environment in Finland. Bulletin of the Geological central Keewatin, Canada. Canadian Journal of Society of Finland, 69: 73–86. Earth Sciences, 15: 1053–1068. Seppälä, M., 1997b. Distribution of permafrost in Sihvo, J., 2001. Ylä-Lapin luonnonhoitoalueen ja Finland. Bulletin of the Geological Society of Finland, Urho Kekkosen kansallispuiston luontokartoi- 69: 87–96. tus. Loppuraportti osa 1: Projektikuvaus. Met- Seppälä, M., 1998. New permafrost formed in peat sähallituksen luonnonsuojelujulkaisuja A130. Met- hummocks (pounus), Finnish Lapland. Perma- sähallitus, Helsinki. 76 pp. frost and Periglacial Processes, 9: 367–373. Sihvo, J., 2002. Ylä-Lapin luonnonhoitoalueen ja Seppälä, M., 2002. Relief control of summer wind Urho Kekkosen kansallispuiston luontokartoi- direction and velocity: a case study from Finn- tus. Loppuraportti osa 2: Ylä-Lapin luototyypit. ish Lapland. Norsk Geografisk Tidsskrift, 56: 117– Metsähallituksen luonnonsuojelujulkaisuja A137. 121. Metsähallitus, Helsinki. 175 pp. Seppälä, M., 2003a. Surface abrasion of palsas by Söderman, G., 1980. Slope processes in cold en- wind action in Finnish Lapland. Geomorphology, vironment of Northern Finland. Fennia, 158: 52: 141–148. 83–152. Seppälä, M., 2003b. An experimental climate change Söderman, G., 1982. Boulder depressions in south- study of the effect of increasing snow cover on ern Finland (in Swedish with English summary). active layer of a palsa, Finnish Lappland. In: Geografisk Tidsskrift, 82: 77–81. Phillips, M., Springman, S.M. & Arenson, L.U. Sokal, R.R. & Rohlf, F.J., 1981. Biometry. The principles Permafrost, Eight international conference, Proceedings, and practice of statistics in biological research. 2nd ed. 1013–1016. A.A. Balkema Publishers, Lisse. W.H. Freeman, New York. 859 pp. Seppälä, M., 2004a. Palsa. In: Goudie, A.S. (ed.). En- Sollid, J.L. & Sørbel, L., 1998. Palsa bogs as a climate cyclopedia of geomorphology, 756–758. Routledge, indicator – examples from Dovrefjell, Southern London & New York. Norway. Ambio, 27: 287–291. Seppälä, M., 2004b. Wind as a geomorphic agent in cold Summerfield, M.A., 1991. Global geomorphology. An climates. Cambridge University Press, Cambridge. introduction to the study of landforms. Longman, Sin- 358 pp. gapore. 537 pp. Seppälä, M., 2005a. Periglacial environment. In: Sep- Svensson, H., 1969. A type of circular lakes in pälä, M. (ed.).The physical geography of Fennoscandia, northernmost Norway. Geografiska Annaler, 51A: 349–364. Oxford University Press, Oxford. 1–12. Seppälä, M., 2005b. Present-day geomorphic proc- Svensson, H., 1971. Some geomorphic features of

145 References

the Varanger Peninsula based on new aerial pho- forms and processes in the timberline region of tographs (in Swedish). Svensk Geografisk Årsbok, northern Finnish Lapland. Zeszyty Naukowe Uni- 47: 7–28. wersytetu Jagiellonskiego, Prace Geograficzne-Zeszyt, Swets, J.A., 1988. Measuring the accuracy of diag- 98: 68–90. nostic systems. Science, 240: 1285–1293. Tricart, L., 1968. Periglacial landscape. In: Fair- Syrilä, S., 1964. Retreat of the continental ice and bridge, R.W. (ed.). The encyclopedia of geomorphol- fluvioglacial erosion features on the Tuolban- ogy, 829–833. Reinhold Book Co., New York. jaugoaivi fjeld in northernmost Finland. Reports Troll, C., 1944. Structurböden, solifluktion und from the Kevo Subarctic Researche Station, 1: 338– frostklimate der Erde. Geologische Rundschau, 34: 345. 545–694. Tabuchi, H. & Hara, Y., 1992. Block fields and Tufnell, L., 1972. Ploughing blocks with special ref- sorted polygons in Finnish Lapland (in Japanese erence to north-west England. Biuletyn Peryglac- with English summary). Geographical Review of Ja- jalny, 21: 237–270. pan, 65: 105–113. Tuhkanen, S., 1980. Climatic parameters and in- Tanner, V., 1915. Studier öfver kvartärsystemet i dices in plant geography. Acta Phytogeographica Fennoskandias nordliga delar III. Fennia, 36. 815 Suecica, 67: 1–110. pp. Tyurin, A.I., 1983. Classification of rock streams. Tanner, V., 1938. Die Oberflächengestaltung In: Miller, K.D. (ed.). Permafrost, Fourth interna- Finnlands. Bidrag till kännedom av Finlands natur tional Conference, Proceedings. 1283–1285. National och folk, 86. 762 pp. Academy Press, Washington D.C. Tarnocai, C. & Zoltai, S.C., 1978. Earth hummocks Ulfstedt, A.-C., 1993. Solifluction in the Swedish of the Canadian Arctic and Subarctic. Arctic and mountains; distribution in relation to vegetation Alpine Research, 10: 581–594. and snow cover. In: Frenzel, B., Matthews, J.A. Thie, J., 1974. Distribution and thawing of perma- & Glaeser, B. (eds.). Solifluction and climate variation frost in the southern part of the discontinuous in the Holocene. Paleoklimaforschung, 11: 217–223. permafrost zone in Manitoba. Arctic, 27: 189– van Everdingen, R., 2005 (ed.). Multi-lan- 200. guage glossary of permafrost and relat- Thorn, C.E., 1979. Ground temperatures and surfi- ed ground-ice terms. 20th of June 2005. cial transport in colluvium during snowpatch meltout: Colorado Front Range. Arctic and Al- Vanhala, H., Suppala, I., Lintinen, P., Hirvas, H. pine Research, 11: 41–52. & Ojala, A.E.K., 2005. Application of electri- Thorn, C.E., 1988. Nivation: a geomorphic chimera. cal and electromagnetic methods for study- In: Clark, M.J. (ed.). Advances in periglacial geomor- ing frozen ground and bedrock – results from phology, 3–31. John Wiley & Sons, Chichester. Ridnitšohkka, Northern Finland. In: Ojala, Thorn, C.E., 1992. Periglacial geomorphology: A.E.K. (ed.). Quaternary studies in the northern and what, where, when? In: Dixon J.C. & Abrahams, Arctic regions of Finland. Geological Survey of Fin- A.D. (eds.). Periglacial geomorphology, 1–30. John land, Special Paper, 40: 13–22. Wiley & Sons, Chichester. Van Houwelingen, J.C. & Le Cessie, S., 1990. Pre- Thorn, C.E. & Hall, K., 1980. Nivation; an Arctic- dictive value of statistical models. Statistics in alpine comparison and reappraisal. Journal of Medicine, 9: 1303–1325. Glaciology, 25: 109–124. van Vliet-Lanoë, B., 1988a. The significance of cry- Thorn, C.E. & Hall, K., 2002. Nivation and cryo- oturbation phenomena in environmental recon- planation; the case for scrutiny and integration. struction. Journal of Quaternary Science, 3: 85–96. Progress in Physical Geography, 26: 533–550. van Vliet-Lanoë, B., 1988b. The origin of patterned Thorn, C.E., Darmody, R.G., Dixon, J.C. & Sch- grounds in Svalbard. In: Senneset, K. (ed.). Per- lyter, P., 2001. The chemical weathering regime mafrost, Fifth International conference, Proceedings, of Kärkevagge, arctic-alpine Sweden. Geomor- 1008–1013. Tapir, Trondheim. phology, 41: 37–52. van Vliet-Lanoë, B., 1991. Differential frost heave, Tikkanen, M. & Heikkinen, O., 1995. Aeolian land- load casting and convection: converging mecha-

146 References

nisms; a discussion of the origin of cryoturba- Mapping, modeling, and visualization of the in- tion. Permafrost and Periglacial Processes, 2: 123– fluences of geomorphic processes on the alpine 139. treeline ecotone, Glacier National Park, MT, van Vliet-Lanoë, B. & Seppälä, M., 2002. Stratig- USA. Geomorphology, 53: 129–145. raphy, age and formation of peaty earth hum- Walters, J.C., 1983. Sorted patterned ground in mocks (pounus), Finnish Lapland. The Holocene, ponds and lakes of the High Valley/Tangle 12: 187–199. Lakes region, central Alaska. In: Miller, K.D. van Vliet-Lanoë, B., Seppälä, M. & Käyhkö, J., 1993. (ed.). Permafrost, Fourth international Conference, Dune dynamics and cryoturbation features con- Proceedings. 1350–1355. National Academy Press, trolled by Holocene water level change, Hieta- Washington D.C. tievat, Finnish Lapland. Geologie en Mijnbouw, 72: Washburn, A.L., 1947. Reconnaissance geology 211–224. of portions of Victoria Island and adjacent re- Venables, W.N. & Ripley, B.D., 2002. Modern applied gions, Arctic Canada. Geological Society of America, statistics with S. 4th ed. Springer–Verlag, New Memoir, 22. 142 pp. York. 495 pp. Washburn, A.L., 1956. A classification of patterned Verstappen, H.Th., 1983. Applied geomorphology. Geo- ground and review of suggested origin. Bulletin morphological survey for environmental developments. of Geological Society of America, 67: 823–866. Elsevier, Amsterdam. 437 pp. Washburn, A.L., 1967. Instrumental observations Virtanen, T., Neuvonen S., & Nikula, A., 1998. of mass-wasting in the Mesters Vig district, Modelling topoclimatic patterns of egg mortali- Northeast Greenland. Meddelelser om Grønland, ty of Epirrita autumnata (Lepidoptera: Geometri- 166. 318 pp. dae) with a Geographical Information System: Washburn, A.L., 1969. Weathering, frost action, and predictions for current climate and warmer sce- patterned ground in the Mesters Vig district, narios. Journal of Applied Ecology, 35: 311–322. Northeast Greenland. Meddelelser om Grønland, Vorren, K.-D., 1967. Evig tele i Norge. Ottar, 51: 176. 303 pp. 1–26. Washburn, A.L., 1970. An approach to a genetic Walker, D.A., Epstein, H.E., Gould, W.A., Kelley, classification of patterned ground. Acta Geo- A.M., Kade, A.N., Knudson, J.A., Krantz, W.B., graphica Lodziensia, 24: 437–446. Michaelson, G., Peterson, R.A., Ping, C-L., Washburn, A.L., 1979. Geocryology - a survey of per- Raynolds, M.K., Romanovsky, V.E. & Shur, Y., iglacial processes and environments. Arnold, London. 2004. Frost-boil ecosystems; complex interac- 406 pp. tions between landforms, soils, vegetation and Washburn, A.L., 1997. Plugs and plug circles; a ba- climate. Permafrost and Periglacial Processes, 15: sic form of patterned ground, Cornwallis Island, 171–188. Arctic Canada; origin and implications. Geological Walsh, C. & MacNally, R., 2003. Hierar- Society of America, Memoir, 190. 87 pp. chical partitioning. R project for sta- Washburn, A.L., 1999. A High Arctic frost-creep/ tistical computing. 20th of April 2004. gelifluction slope, 1981–89: Resolute Bay, Corn- wallis Island, Northwest Territories, Canada. Walsh, S.J., Butler, D.R. & Malanson, G.P., 1998. An Permafrost and Periglacial Processes, 10: 163–186. overview of scale, pattern, process relationships Watson, E., 1966. Two nivation cirques near Aber- in geomorphology: a remote sensing and GIS ystwyth, Wales. Biuletyn Peryglacjalny, 15: 79–101. perspective. Geomorphology, 21: 183–205. Werner, B.T. & Hallet, B., 1993. Numerical simula- Walsh, S.J., Bian, L., McKnight, S., Brown, D.G. & tion of self-organized stone stripes. Nature, 361: Hammer, E.S., 2003a. Solifluction steps and ris- 142–145. ers, Lee Ridge, Glacier National Park, Montana, Wilcock, P.R. & Iverson, R.M., 2003 (eds.). Prediction USA; a scale and pattern analysis. Geomorphology, in geomorphology. American Geophysical Union, 55: 381–398. Washington, DC. 256 pp. Walsh, S.J., Butler, D.R., Malanson, G.P., Crews- Williams, P.J., 1959. The development and signifi- Meyer, K.A., Messina, J.P. & Xiao, N., 2003b. cance of stony earth circles. Skrifter Utgitt an det

147 References

Norsk Videnskaps Akademi i Oslo, Matematisk- Zevenbergen, L.W. & Thorne, C.R., 1987. Quanti- Naturhistorisk Klasse, 3. 14 pp. tative analysis of landsurface topography. Earth Williams, P.J. & Smith, M.W., 1989. The frozen earth. Surface Processes and Landforms, 12: 47–56. Fundamentals of geocryology. Cambridge University Zhang, J. & Goodchild, M.F., 2002. Uncertainty in geo- Press, Cambridge. 306 pp. graphical information. Taylor & Francis, London. Wilson, L., 1968. Morphogenetic classification. 266 pp. In: Fairbridge, R.W. (ed.). The encyclopedia of geo- Zimmermann, N., 2003. 1.1 shortwave.aml. 15th morphology, 717–729. Reinhold Book Co., New of April 2003. York. analysis. In: Wilson, J.P. & Gallant, J.C. (eds.). Zuidhoff, F.S., 2002. Recent decay of a single palsa Terrain Analysis. Principles and Applications, 1–27. in relation to weather conditions between 1996 John Wiley & sons, New York. and 2000 in Laivadalen, northern Sweden. Ge- Wilson, P. & Sellier, D., 1995. Active patterned ografiska Annaler, 84A: 103–111. ground and cryoturbation on Muckish Moun- Zuidhoff, F.S. & Kolstrup, E., 2000. Changes in tain, Co. Donegal, Ireland. Permafrost and Perigla- palsa distribution in relation to climate change cial Processes, 6: 15–25. in Laivadalen, Northern Sweden, especially Worsley, P., 2004. Periglacial geomorphology. In: 1960–1997. Permafrost and Periglacial Processes, 11: Goudie, A.S. (ed). Encyclopedia of geomorphology, 55–69. 772–776. Routledge, London & New York. Zuidhoff, F.S. & Kolstrup, E., 2005. Palsa devel- Yamada, S., Matsumoto, H. & Hirakawa, K., 2000. opment and associated vegetation in northern Seasonal variation in creep and temperature in a Sweden. Arctic, Antarctic and Alpine research, 37: solifluction lobe; continuous monitoring in the 49–60. Daisetsu Mountains, northern Japan. Permafrost Zweig, M.H. & Campbell, G., 1993. Receiver-oper- and Periglacial Processes, 11: 125–135. ating characteristic (ROC) plots: a fundamental Yee, T.W. & Mitchell, N.D., 1991. Generalized addi- evaluation tool in clinical medicin. Clinical Chem- tive models in plant ecology. Journal of Vegetation istry, 39: 561–577. Science, 2: 587-602.

148 Other source material

OTHER SOURCE MATERIAL

Black-and-white aerial photographs 1:31 000, 1961. 61170 241–250, 268–279, 283–293, 372–381 and 438–456. National Land Survey of Finland, Helsinki. Black-and-white aerial photographs 1:31 000, 1961. 61182 227–235. National Land Survey of Fin- land, Helsinki. Black-and-white aerial photographs 1:31 000, 2001. 0111 31, 62 and 64. National Land Survey of Finland, Helsinki. NSL Topographic Database, 2000. Contours, lakes and ponds. National Land Survey of Finland, Helsinki. Digital black-and-white aerial photographs, 2000. OrtoCD. National Land Survey of Finland, Helsinki. Topographical maps 1:20 000, 1975. Sheets 3913 01, 02 and 04–06. National Land Survey of Fin- land, Helsinki. Topographical map 1:20 000, 1977. Sheet 3913 03 + 3911 12. National Land Survey of Finland, Helsinki.

149