Modeling and Forecasting the Influence of Current and Future

The University of Maine DigitalCommons@UMaine

Electronic Theses and Dissertations Fogler Library

Winter 12-2016 Modeling and Forecasting the Influence of Current and Future Climate on Eastern North American Spruce-Fir (Picea-Abies) Forests Caitlin Andrews [email protected]

Follow this and additional works at: http://digitalcommons.library.umaine.edu/etd Part of the Forest Management Commons, and the Other Forestry and Forest Sciences Commons

Recommended Citation Andrews, Caitlin, "Modeling and Forecasting the Influence of Current and Future Climate on Eastern North American Spruce-Fir (Picea-Abies) Forests" (2016). Electronic Theses and Dissertations. 2562. http://digitalcommons.library.umaine.edu/etd/2562

This Open-Access Thesis is brought to you for free and open access by DigitalCommons@UMaine. It has been accepted for inclusion in Electronic Theses and Dissertations by an authorized administrator of DigitalCommons@UMaine. MODELING AND FORECASTING THE INFLUENCE OF CURRENT AND FUTURE CLIMATE ON EASTERN NORTH AMERICAN SPRUCE-FIR (PICEA-ABIES) FORESTS

Caitlin Andrews

B.S. University of Vermont, 2009

A THESIS Submitted in Partial Fulfillment of the Requirements for the Degree of Master of Science (in Forest Resources)

The Graduate School

University of Maine

December 2016

Advisory Committee:

Aaron Weiskittel, Associate Professor of Forest Biometric and Modeling, Advisor

Anthony D’Amato, Associate Professor in Silviculture and Applied Forest Ecology

Erin Simons-Legaard, Assistant Research Professor in Forest Landscape Modeling

THESIS ACCEPTANCE STATEMENT

On behalf of the Graduate Committee for Caitlin M Andrews I affirm that this manuscript is the final and accepted thesis. Signatures of all committee members are on file with the Graduate School at the University of Maine, 42 Stodder Hall, Orono, Maine.

Aaron Weiskittel, Associate Professor of Forest Biometrics and Modeling 12/9/2016

iii

LIBRARY RIGHTS STATEMENT

In presenting this thesis in partial fulfillment of the requirements for an advanced degree at the University of Maine, I agree that the Library shall make it freely available for inspection. I further agree that permission for “fair use” copying of this thesis for scholarly purposes may be granted by the Librarian. It is understood that any copying or publication of this thesis for financial gain shall not be allowed without my written permission.

Signature:

Date: December 9, 2016

MODELING AND FORECASTING THE INFLUENCE OF CURRENT AND FUTURE CLIMATE ON EASTERN NORTH AMERICAN SPRUCE-FIR (PICEA-ABIES) FORESTS

By Caitlin Marie Andrews

Thesis Advisor: Dr. Aaron R. Weiskittel

An Abstract of the Thesis Presented in Partial Fulfillment of the Requirements for the Degree of Master of Science (in Forest Resources) December 2016

The spruce-fir (Picea-Abies) forest type of the Acadian Region is at risk of disappearing from the United States and parts of Canada due to climate change and associated impacts. Managing for the ecosystem services provided by this forest type requires accurate forecasting of forest metrics across this broad international region in the face of the expected redistribution of tree species. This analysis linked species specific data with climate and topographic variables using the nonparametric random forest algorithm, to generate models that accurately predicted changes in species distribution due to climate change. A comprehensive dataset, consisting of 10,493,619 observations from twenty-two agencies, including historical inventories, assured accurate assignation of species distribution at a finer resolution (1 km2) than previous analyses. Different dependent variables were utilized, including presence/absence, a likelihood value, abundance variables

(i.e. basal area, stem density, and importance value), and predicted maximum stand density index (SDImax), in order to inspect the difference in results in regards to their conservation management utility, as well as the effects of inherent species life history traits on outcomes.

Using linear quantile mixed models, predictions of SDImax were estimated for spruce or fir-dominated plots across the Acadian Region. Model performance was strong and estimates of SDImax from these models were similar to previous regional studies. The establishment of an individual constant slope of self-thinning for plots dominated by each spruce or fir species reinforces previous research that Reineke’s slope is not universal for all species, and that the differences in slope are telling of different species’ life history patterns. Individual plot estimates of SDImax, achieved through a varying intercept, allowed for the assessment of each stand’s potential and limitations in regards to the impact that climate, nutrient availability, site quality, and other factors might have on SDI.

A high association with environmental variables was exhibited for all dependent variables. Area under receiver operator curve values for presence/absence models averaged

0.99 ± 0.01 (mean ± SD) well above the accepted standard for excellent model performance.

The addition of historical tree data revealed supplementary suitable habitat along the southern edge of species’ ranges, due to marginal dynamics potentially overlooked by approaches relying solely on current inventories. The likelihood models provided an adequate surrogate to abundance models, reflecting gradients of suitable habitat. The

SDImax variables performed the best of the continuous variables inspected in regards to climate associations, likely because of the selection of spruce or fir-dominated plots and the ability to capture core ranges. Black spruce (Picea mariana (Miller) B.S.P.) responded the best to abundance modeling, due to this species’ uniform range. White spruce (Picea glauca

(Moench) Voss) consistently performed the worst among all species for each model, due to this species’ wide distribution at low abundances. Presence/absence models assist in understanding the full range of climatically suitable habitats, abundance values provide the

ability to prioritize suitable habitat based upon higher abundance, and SDImax models can be utilized for the construction of Density Management Diagrams and the active management of future landscapes based on size-density relationships.

ACKNOWLEDGEMENTS

It would not have been possible to write this thesis without the help and support of many of the amazing people around me, to only some of whom it is possible to mention here.

This thesis would not have been possible without the persistent guidance and patience of my principal advisor, Dr. Aaron Weiskittel, whose knowledge of forest statistics and modeling is inspiring. I am eternally grateful for exposure to this field that has transformed my career ambitions. I would also like to thank my committee members, Dr.

Anthony D’Amato and Dr. Erin Simons-Legaard, who I am extremely appreciative of for their assistance and suggestions throughout my thesis on topics of forest management and modeling.

This project would not have been possible without the collaboration of multiple agencies, and to the individuals who helped collect, organize, and distribute the data necessary for this research, I am beholden. In particular I would like to thank Ingeborg

Seaboyer of the Caroline A. Fox Research and Demonstration Center, Shawn Meng of the

Manitoba Forestry Branch, Thom Snowman of the Massachusetts Department of

Conservation and Recreation, Daniel Wovcha of the Minnesota Department of Natural

Resources, Rebecca Key, Kathyrn Miller, Suzanna Sanders, and Scott Weyenberg of the

National Park Service, Jennifer Weimar at the New Hampshire Division of Forest of Lands,

John Parton of the Ontario Ministry of Natural Resources, Dr. Shawn Fraver of the University of Maine, William DeLuca of the University of Massachusetts, Judith Scarl of the Vermont

Center for Ecostudies, and James Duncan of the Vermont Monitoring Cooperative for

iv assisting in the acquisition of data. In particular, I would like to thank Charles Cogbill, for his independent and arduous collection of historical data throughout New England, and for sharing this data and insights on historical forest composition throughout the Region.

I would like to acknowledge the School of Forest Resources at the University of

Maine for providing academic and technical support for my thesis, and for inspiring an everlasting love and respect for the world of forestry. Additionally, I would to thank NASA for their continued financial support throughout my project, which allowed for the exploration of innumerable facets of my thesis topic.

Lastly, I am most thankful to my fiancé Jorge, for his bottomless patience and continued faith in my ability to succeed, and to my pets, Faith, Justice, Boquillas, and Alsate, who provide comfort in good times and bad. I wouldn’t be here without it.

TABLE OF CONTENTS

ACKNOWLEDGEMENTS ...... iv

LIST OF TABLES ...... x

LIST OF FIGURES ...... xii

CHAPTER 1 INTRODUCTION AND REVIEW OF THE SPRUCE-FIR FOREST AND SPECIES-

CLIMATE MODELING ...... 1

1.1. Introduction ...... 1

1.2. The Acadian Forest ...... 3

1.3. Species-Climate Associations ...... 7

1.4. Statistical and Mechanistic Models ...... 11

1.5. The Dependent Variable ...... 13

1.6. Objectives ...... 18

CHAPTER 2 MODELING AND FORECASTING EASTERN NORTH AMERICAN SPRUCE-FIR

OCCURRENCE/ABUNDANCE UNDER CURRENT AND FUTURE CLIMATE

CONDITIONS ...... 19

2.1. Abstract ...... 19

2.2. Introduction ...... 20

2.3. Methods ...... 27

2.3.1. Study Area ...... 27

2.3.2. Tree Data...... 28

2.3.2.1. Contemporary Tree Data ...... 29

2.3.2.2. Historical Tree Data...... 30

2.3.3. Climate Data ...... 30

2.3.4. Topographic Data ...... 31

2.3.5. Species-Specific Distribution Model Development ...... 34

2.3.6. Model Evaluation and Comparison...... 37

2.3.7. Predictive Mapping ...... 40

2.4. Results ...... 40

2.4.1. Data Characteristics ...... 40

2.4.2. Model Performance ...... 41

2.4.3. Historical Model Performance ...... 50

2.4.4. Likelihood Model Performance ...... 52

2.4.5. Future Predictions of Species’ Distributions ...... 53

2.5. Discussion ...... 60

2.6. Conclusion ...... 73

CHAPTER 3 MODELING AND FORECASTING THE INFLUENCE OF CURRENT AND FUTURE

CLIMATE ON MAXIMUM STAND DENSITY FOR EASTERN NORTH AMERICAN

SPRUCE-FIR FORESTS ...... 74

3.1. Abstract ...... 74

3.2. Introduction ...... 76

3.3. Methods ...... 83

3.3.1. Data ...... 83

vii

3.3.1.1. Climate Data ...... 84

3.3.1.2. Topographic Data ...... 85

3.3.2. LQMM Analysis ...... 85

3.3.3. Random Forest Analysis ...... 90

3.3.4. Current and Future Predictions ...... 92

3.4. Results ...... 93

3.5. Discussion ...... 103

3.6. Conclusion ...... 114

CHAPTER 4 CONCLUSION AND SUMMARY OF MAJOR FINDINGS ...... 116

4.1. Summary of Findings by Objective ...... 117

4.1.1. To Explore New Data and Modeling Techniques for SDMs ...... 117

4.1.2. To Characterize the Distribution and Abundance of the Important Species in

the Spruce-Fir Forest, while Comparing the Usefulness of Both

Presence/Absence and Abundance Models, as well as Alternatives, for

Conservation Decisions ...... 119

4.1.3. To Compare and Illustrate the Differences between the Results and Application

of Directly Calculated Variables Useful for Passive Management versus

Predicted Variables Useful for the Active Management of Forests ...... 121

4.2. Summary of Findings by Species ...... 124

4.2.1. Balsam Fir (Abies balsamea L.) ...... 124

4.2.2. White Spruce (Picea glauca (Moench) Voss) ...... 125

viii

4.2.3. Black Spruce (Picea mariana (Miller) B.S.P.) ...... 126

4.2.4. Red Spruce (Picea rubens Sarg.) ...... 127

4.3. Conclusion ...... 128

BIBLIOGRAPHY ...... 130

APPENDICES ...... 150

APPENDIX A: Data Sources ...... 151

APPENDIX B: Effect of Tree Diameter Thresholds on Analysis ...... 159

APPENDIX C: Effect of Solely Using Forest Inventory and Analysis Data for Acadian

Forest Spruce-Fir Species Distribution Models ...... 164

APPENDIX D: Testing the Output of Likelihood Models as a Predictor of Abundance ...... 167

BIOGRAPHY OF THE AUTHOR ...... 174

LIST OF TABLES

Table 2.1. Description of climate variables used in analysis… ...... 32

Table 2.2. Statistics of abundance values by species...... 43

Table 2.3. Results of random forest analysis for each species...... 45

Table 2.4. Kappa values for all models...... 58

Table 2.5. Area (thousands of kilometers) occupied by each species under three different

models...... 62

Table 3.1. Description of climate variables used in this analysis...... 86

Table 3.2. Summary of stand variables for each species...... 94

Table 3.3. Statistics of predicted intercept and slope for each species’ linear quantile

mixed model ...... 96

Table 3.4. Summary of predicted variables for each species using linear quantile mixed

models...... 99

Table 3.5. Summary of stand variables for each species in the 99th percentile...... 99

Table 3.6. Results of the SDImax random forest models for each species...... 99

Table 3.7. Summary of SDImax predicted using the random forest models for each

species...... 101

Table A.1. Description of different data sources used in analyses...... 156

Table B.1. Results of random forest analyses for presence/absence modeling

performed with a threshold of 1 cm and 5 cm as a requirement for

individuals included in analysis...... 159

Table C.1. Results for presence/absence modeling with only US Forest Service Forest

Inventory and Analysis data...... 165

Table D.1. Values for negative binomial model comparison models...... 170

Table D.2. Coefficients for zero-inflated model (ZIM) for each species...... 171

Table D.3. Observed versus predicted frequencies for negative binomial models for

each species...... 172

LIST OF FIGURES

Figure 1.1. Map of the Acadian Region...... 4

Figure 2.1. Study area overlaid with World Wildlife Fund (WWF) ecoregions...... 27

Figure 2.2. Frequency of types of data sorted by ecoregion...... 43

Figure 2.3. Density plots for actual versus predicted basal area (a), relative stem

density (b), and importance value (c) per species...... 47

Figure 2.4. Presence versus absence plots' relationship with PRMTCM per species...... 49

Figure 2.5. Actual and predicted presence for each species...... 51

Figure 2.6. Actual and predicted stem density for each species...... 54

Figure 2.7. Actual and predicted importance value (IV) for each species...... 55

Figure 2.8. Actual and predicted basal area (BA) and likelihood model outputs for each

species ...... 56

Figure 2.9. Boxplots exhibiting the relationship between predicted likelihood and

predicted relative area abundance for each species...... 59

Figure 2.10. Future predicted presence or absence for each species...... 63

Figure 2.11. Future predicted likelihood for each species...... 64

Figure 3.1. Map of different spruce-fir forest types distributed across the study area...... 94

Figure 3.2. Observed ln (TPH) vs observed ln (DR) for all plots used in this analysis...... 95

Figure 3.3. Predicted ln (TPH) vs. observed ln (DR) for all plots in this analysis...... 97

Figure 3.4. Maps of current actual versus predicted SDImax...... 100

Figure 3.5. Partial dependency plots for each species’ random forest model and the

two most important variables from those models...... 102

xii

Figure 3.6. Maps of future predictions of SDImax depicted as a ratio between the

future predicted value and the current predicted value...... 104

Figure B.1. Mapped predictions of presence/absence models using a data inclusion

threshold of 1 cm and 5 cm for balsam fir...... 160

Figure B.2. Mapped predictions of presence/absence models using a data inclusion

threshold of 1 cm and 5 cm for white spruce...... 161

Figure B.3. Mapped predictions of presence/absence models using a data inclusion

threshold of 1 cm and 5 cm for black spruce...... 162

Figure B.4. Mapped predictions of presence/absence models using a data inclusion

threshold of 1 cm and 5 cm for red spruce...... 163

Figure C.1. Mapped predictions objects for presence/absence models for each species

generated with solely United States Forest Service Forest Inventory and

Analysis data...... 166

xiii

CHAPTER 1

INTRODUCTION AND REVIEW OF THE SPRUCE-FIR FOREST AND SPECIES-CLIMATE

MODELING

1.1. Introduction

It is certain that global surface temperatures have increased since measurement began in the late 19th century (Stocker et al., 2013). Temperatures on average have risen 0.89°C since 1880, with 80% of the increase occurring after 1950. Furthermore, climate models predict with high confidence that the 30-year period between 1982 and 2012 is the warmest 30-year period of the last 800 years. This increase in temperatures has cascading effects on sea surface temperatures, annual precipitation, glacier and ice sheet volume, and many more aspects of the global climate system. These changes to climate are unsurprisingly reflected in species’ distributions and ecosystems’ configurations. It is recognized that as temperatures rise species’ geographic distributions generally shift poleward and upward in altitude (Harsch et al., 2009; Lenoir et al., 2008; Parmesan, 2006).

Paleoecological evidence confirms that temperature shifts as little as 1°C led to significant forest reconfigurations as little as 1,000 years ago (Lindbladh et al., 2003; Schauffler and

Jacobson, 2002). Currently, transformations are already being witnessed, with one meta- analysis of mobile organisms estimating a median latitudinal migration of 16.9 km per decade and a median shift to higher elevations of 11 m per decade (Chen et al., 2011).

Climate impacts on sessile flora, such as forests, are still being evaluated, as response to climate change is complex, relying on the interactive effects of both temperature and precipitation changes (Parmesan, 2006). Rapid migration potential is limited, and shifts in

1 the suitability of habitat conditions (Iverson et al., 2008), or the reconfiguration of forest structure, composition, and productivity (Dolanc et al., 2013; Mohan et al., 2009), are a common outcome of climate warming.

According to the latest report by the Intergovernmental Panel on Climate Change

(IPCC), it is extremely likely that more than half of the observed increases in global temperatures can be assigned to anthropogenic influences, including greenhouse gas emissions and land use changes (Rosenzweig et al., 2008; Stocker et al., 2013). Future projections of climate are based upon our knowledge of anthropogenic and natural influences to the system, as well as scenarios based upon how humans may or may not mitigate climate change over the next century. Assuming sustained doubling of atmospheric carbon dioxide (CO2), models indicate that temperatures will rise between 1.5°C and 4.5°C by 2090, and that a rise less than 1°C or greater than 6°C is extremely unlikely. Feedback effects due to climate change will create regional differences in cloud cover, precipitation, and extreme weather events, necessitating the inspection of localized downscaled models of climate projections. Of particular concern are extreme events, including severe storms

(i.e. hurricanes, northeaster) and extended periods of drought and freezing temperatures, which directly contribute to mass forest mortality, as well as indirectly, through increased vulnerability to wildfire and insect attacks (Allen et al., 2010; Huntington et al., 2009).

Change in climate is already being manifested in the regional redistribution of forests.

Numerous studies have documented the shift of forest habitat (Beckage et al., 2008; Kelly and Goulden, 2008; Lenoir et al., 2008) upward in altitude, or the loss of ecosystems altogether (Condit et al., 1996), due to climate change. Other studies have observed the redistribution of forest structure as a result of the mortality of mature individuals (Dolanc et

2 al., 2013). In general, climate effects to forest ecosystems are either chronic, through gradual changes in the central tendencies of climatic variables (Adams et al., 2009; Beckage et al., 2008) or abrupt (Shuman et al., 2009), including extreme events such as drought in water stressed ecosystems (Park Williams et al., 2012) or rising sea-level in tidal ecosystems

(Doyle et al., 2010). Evidence of climate related drought and heat stress induced mortality in forest is present on all six of the treed continents (Allen et al., 2010). Warmer temperatures, independent of precipitation amount, can increase forest water stress and shorten the time to drought-induced mortality (Adams et al., 2009; Park Williams et al., 2012). Drought increases vulnerability to additional stressors including wildfire and disease outbreak

(Huntington et al., 2009; Noss, 2001). Observed increases in the area of forests burned in

Canada over the last four decades is consistent with models due to anthropogenic climate change (Gillett et al., 2004) and all aspects of insect outbreak cycles have intensified as the climate warms (Logan et al., 2003). Not all effects of climate change are adverse, and greater levels of CO2, as well as simultaneous increases in temperature and precipitation, have boosted forest productivity in many locations (Huntington et al., 2009; Parmesan,

2006; Swetnam and Betancourt, 1997). The myriad effects of a changing climate on forest growth and distribution necessitates the inspection of individual ecosystems to properly analyze and predict specific transformations.

1.2. The Acadian Forest

Traversing international boundaries, the Acadian Forest stretches from the northern

New England states of the United States (U.S.) to Québec and the maritime provinces of

Canada (Figure 1.1), and is of great ecological and economic value to the region. Bounded by the boreal forest to the north and the temperate, deciduous hardwood forest to the

3 south, the Acadian Forest is distinct for its mixed-wood stands at higher elevations and the economically important spruce-fir (Picea-Abies) forest type present on lower slopes (Loo and Ives, 2003; Westveld, 1931). The Acadian Forest contains fourteen species of conifers, more than any other mixed forest save the Appalachian Blue Ridge and Southeastern mixed forests, and 35 species of hardwoods (Olson et al., 2001). Of the 49 common tree species,

49% (twenty-three) exhibit a range boundary in the Acadian Region (Barton et al., 2012).

The rich composition of this forest is inextricably linked to the varied climate and it is clear that changes in climate will have effects on forest make-up, as well as the people and wildlife communities that rely on it.

Figure 1.1. Map of the Acadian Region. The dark green represents the Acadian Forest Region designated by the World Wildlife Fund (WWF).

The Acadian Region is expected to have hotter summers with less precipitation and shorter winters marked by more rain and less snow (Jacobson et al., 2009). Projected future changes are consistent with a warmer climate, including shrinking snow cover, more frequent droughts, and extended periods of low hydrological flows in the summer (Hayhoe et al., 2007). Summertime precipitation is projected to decrease on the Acadian coastline and inland, but increase along the Canadian border (Anderson et al., 2010; Hayhoe et al.,

2008). Meanwhile, evaporation is expected to increase in most of the region, resulting in lower soil moisture content and higher humidity (Anderson et al., 2010). Extreme precipitation events are projected to increase by at least 50%, while days with extreme high temperatures are expected to at least double (Anderson et al., 2010; Hayhoe et al., 2008).

Short- and medium-term droughts are expected to increase, and in conjunction with drier hotter summers, the effects on the water supply could be severe (Hayhoe et al., 2007).

Already, overall average temperatures increased by 0.37 to 0.43°C per decade between

1965 and 2005, with greater temperature increases in the winter (Huntington et al., 2009).

The amount of days with snow on the ground has decreased by up to 25 days and ice-out on rivers and lakes has decreased by nine days (Hodgkins et al., 2002; Wake et al., 2006).

This diversity in climate conditions for the Acadian Region can partially be attributed to a correspondingly diverse geography. This region is approximately 23,750,190 ha and spans seven degrees of latitude (Olson et al., 2001). The presence of a long coastline, buffered by the Labrador Current, translates to cooler and moister climatic trends for this area. The southern edge of the Labrador current converges with the much warmer Gulf stream, resulting in a dramatic sea surface temperature shifts and increased atmospheric activity at this boundary (Bradbury et al., 2002). Climate in the region is predominantly controlled by

5 clashing atmospheric circulation patterns that currently convene in the mid-latitudes.

Warm, wet subtropical systems meet sub-polar maritime systems and dry, cold continental arctic masses at the Polar Jet Front. Much of the Acadian Region lies on the boundary of the ever-shifting polar front. While the polar cell typically dips further south in the winter and the Hadley cell pushes further north in the summer, the region can be on either side of the boundary at any time of the year (Keim, 1998; Zielinski and Keim, 2003). Climate predictions are consistent with a summertime northward shift in the Polar Jet Front, resulting in warmer summertime temperatures, and an eastward shift of the East Coast Trough, resulting in drier conditions (Hayhoe et al., 2007).

The Acadian Forest is composed of a complex variety of different forest types, including numerous spruce-fir communities. Within the Acadian Forest, the spruce-fir forest type is a distinguishing feature that provides forest products and wildlife habitat. Spruce-fir communities compose approximately 42% of Canada’s Acadian Forest and 32%, 10%, and

14% of New Hampshire, Maine, and Vermont, respectively, in the U.S. (Canada’s National

Forest Inventory, 2006; McWilliams et al., 2005; North East State Foresters Association,

2007). The forest product industry is led by softwood production due to the availability of this resource. Forest products account for up to 4.9% (Maine) in the USA and 9% (New

Brunswick) in Canada of regional gross domestic products (APEC, 2005, 2003;

Forest2Market, 2009). Several species of local (e.g. spruce grouse (Dendragapus canadensis canace)) and national concern (e.g., Bicknell’s Thrush (Catharus bicknelli), Canadian Lynx

(Lynx canadensis)) rely on the spruce-fir forest for habitat.

Traditionally, Acadian spruce-fir forests were broadly divided into two types: dominant softwood and secondary softwood. Dominant softwood includes spruce swamps, spruce-fir

6 flats, high elevation spruce slopes, and the coastal spruce-fir. Secondary softwoods include yellow birch (Betula alleghaniensis Britton)-spruce and sugar maple (Acer saccharum

Marsh)-spruce forest types (Hosmer, 1902; Leak, 1982; Mosseler et al., 2003). While human disturbance has undeniably altered the landscape and distorted forest types, these spruce- fir forests are still recognizable today. Recent surveys have similarly grouped different spruce-fir types, but with more detail. The United States Forest Service (USFS) Forest

Inventory and Analysis (FIA) program makes use the Society of American Foresters’ (SAF) classification system, which lists six different spruce-fir types for the Acadian Region (Eyre,

1980). One recent classification only for Maine includes ten different community types with a majority spruce-fir component. These include black spruce barrens, black spruce woodlands, lower elevation spruce-fir forests, maritime spruce-fir forests, spruce rocky woodlands, montane spruce-fir forests, subalpine fir forests, spruce-pine woodlands, spruce-northern hardwoods and black spruce bogs (Gawler and Cutko, 2010). It is evident that spruce-fir forest assemblages are diverse and that when referring to this forest type we are talking about a spectrum of geographic, edaphic, and climatic conditions.

1.3. Species-Climate Associations

The Acadian spruce-fir forest type relies on cooler and moister conditions associated with northern latitudes and sensitive high alpine and coastal areas, and is at a particular risk for loss of habitat due to climate change. Previous climate models have predicted range contraction of up to 400 kilometers north (Iverson et al., 2008) and a possible reduction of

97-100% of suitable habitat in the U.S. in the next 100 years (Hansen et al., 2001). Refugia locations in New England are predicted to be restricted to high elevations or inland along the U.S.-Canada border (Tang and Beckage, 2010). These studies of the spruce-fir forest

7 have been limited by the absence of data that fully characterizes the species’ relationships with the environment in the northern portion of their ranges, as they reach across international boundaries. The absence of this data not only limits understanding of the species relationship with climate, but also prohibits recognizing future suitable habitat for forest communities.

In order to better understand the predictions of species’ distributions, and to envision how future landscapes might manifest themselves, understanding individual species’ physiological tolerances and optima in regards to not only range boundaries, but also life history requirements, is essential. Recent biogeographical studies suggest that tolerance to climate extremes, particularly freezing temperatures, accounts for 80% of variation in range size (Mathews and Bonser, 2005; Pither, 2003). Since recent climate trends are particularly driven by warming winter temperatures (Stocker et al., 2013), the assumption is that tree species’ ranges currently restricted by freezing temperatures will expand or experience increased growth at the edges of their ranges (Harsch et al., 2009). On the other hand, soil moisture is critical to seedling recruitment success (Chmura et al., 2011; Greenwood et al.,

2008), and as temperatures warm, not only is soil moisture predicted to decrease (Anderson et al., 2010), but longer, more frequent episodes of drought are expected (Hayhoe et al.,

2008). Additionally, it is important to recognize the impact of biotic interactions on species’ ranges, as this certainly influences the realized niche witnessed on the current landscape and is often a result of physiological limitations in regards to light tolerance, rooting depth, and nutrient requirements in the face of competition (Schwarz et al., 2003). As climate changes realized niches will shift within the bounds of their fundamental niche (Maiorano et al., 2013), and phenotypic variation will be expressed as a response to changing conditions

(Kearney and Porter, 2009). The primary species of the Acadian spruce-fir forest types are balsam fir (Abies balsamea L.), white spruce (Picea glauca (Moench) Voss), black spruce

(Picea mariana (Miller) B.S.P.), and red spruce (Picea rubens Sarg.). While these species exist in distinct associations with one another today, paleoecology studies indicate that past compositions have no bearing on current, and likely future, forest assemblages (Davis, 1976;

Huntley, 1991).

Black and white spruce are thought of as “plastic” species, meaning they can survive in highly variable circumstances, with extreme climate and soil conditions, and are associated with establishment post-glaciation (Halliday and Brown, 1943; Lindbladh et al., 2003). For example, black spruce was found to survive in one study area where temperatures dipped to -62°C, and white spruce to -54°C (Maini, 1966; Major et al., 2003). Generally, plastic species’ ranges are larger than those with more specific niches (Morin and Lechowicz,

2013), and abundance and frequency of these species within their range are controlled less by abiotic factors, and more by biotic competition (Murphy et al., 2006). Black spruce is more cold tolerant than white spruce, and enjoys near 100% abundance in the core of its range (Vincent, 1965). In the Acadian Region, black spruce’s shallow root system allows for survival in organic and water logged soils including peatlands throughout Canada (Brumelis and Carleton, 1988) and the species will grow in the understory on rich sites due to an intermediate shade tolerance (Vincent, 1965). Black spruce is also much more tolerant of frequent fire, and associated dry weather, than other associated spruce species (Foster,

1983). In eastern North America, white spruce is not nearly as abundant, likely due to the fact that it is more demanding of light and soil conditions than associated conifers

(Kabzems, 1971; Sutton, 1969). Paleoecological reconstructions suggests that white spruce

9 was the first to arrive in post-glacial periods and thrived on rich, coarse-textured soils with good drainage (Lindbladh et al., 2007), but was quickly replaced on the landscape by black spruce due to paludification as the climate became colder and wetter (Grimm and Jacobson,

2003; Lindbladh et al., 2007). White spruce establishes and grows well on abandoned farmland and other select coastal sites due to fast establishment with light availability, though it is outcompeted over time (Davis, 1966).

In the Acadian Region, often suitable habitat for black spruce gives way to genetically and morphologically similar red spruce (Gordon, 1976). Red spruce occupies a much more specific niche than the other spruces of the region, and this is thought to be mostly controlled by adequate moisture in cool environs (Dumais and Prévost, 2007).

Paleoecological evidence suggests that red spruce growth is prohibited in dry warm conditions and is also limited by low winter temperatures (i.e. -16°C, Thompson et al.,

2009), and that the proliferation of this species in New England is a recent phenomenon due to cooler and moister conditions (Lindbladh et al., 2003). Maximum development is obtained at the southern edge of its range, in the humid southern Appalachian mountains

(Walter, 1967), and foggy, coastal habitat in the northeast (Davis, 1966). Adequate moisture is essential for germination (Frank and Bjorkbom, 1973), as well as a mineral soil layer reachable by red spruce’s shallow rooting system (Hart, 1965). Similar to black spruce, red spruce will grow on thin, unformed soils that other species will not tolerate, most notably at high elevations in New England (Frank and Bjorkbom, 1973; Seymour, 1995), though this species is much more frost intolerant than black spruce (Major et al., 2003). Red spruce is very shade tolerant and long-lived, and will persist in the understory for many years as advanced regeneration before being released (Davis, 1991; Seymour, 1992).

Lastly, though not considered a plastic species, nor as cold tolerant as black and white spruce, balsam fir is a generalist with the ability to survive in a wide array of climate and soil conditions. Balsam fir is extremely competitive and flowers and thrives in full light, taking advantage of disturbed environments to establish itself (Bakuzis and Hansen, 1965). Balsam fir is widely believed to have increased in abundance across the landscape due to frequent clear cuts over the last century, particularly after the spruce budworm infestation of the late

1970s (McWilliams et al., 2005). Though the root system of this species is relatively shallow, it is deeper than that of all spruces, spreading faster and deeper during establishment

(Bakuzis and Hansen, 1965; Greenwood et al., 2008), giving it a competitive edge. And while light is an important factor for growth, soil moisture is the most important factor determining seedling establishment, though it is able to succeed in a variety of situations.

1.4. Statistical and Mechanistic Models

Describing the relationship between an ecosystem and its environment as it relates to climate change is typically achieved in one of two ways. One, the ecosystem is examined through the lens of its important species, and a bioclimatic envelope is developed for each species through direct statistical linkages. Also known as species distribution models

(SDMs), ecological niche models, and bioclimatic envelopes, this method is an empirical based approach to correlating the presence of species to climatic variables, assuming the hypothesis that the best indicator of a species realized niche is its current distribution

(Pearson and Dawson, 2003). Direct statistical linkages between environmental variables and species distributions are relatively easily accounted for and evaluated (Araújo et al.,

2005), and the field profits from a long history of use, discussion, and development

(Heikkinen et al., 2006; Luoto et al., 2005). Until recently statistical methods were seen as a

11 poor choice for species-climate modeling as this relationship was hard to capture, but the advent of computer based classification and regression trees (CARTs) has been able to accurately predict associations (Cutler et al., 2007). Obvious limitations for this methodology include the inability to capture the fundamental niche of species, as well as biotic interactions between organisms (Williams et al., 2013). Additionally, extrapolating these models to unknown scenarios, such as future climate change, does not account for species’ genetic variability, phenotypic plasticity, evolutionary changes, CO2 effects, and dispersal pathways (Elith and Leathwick, 2009; Heikkinen et al., 2006). Lastly, studies often suffer from a lack of high quality empirical data that is necessary for accurate predictions.

Alternatively, ecosystems are modeled though prefabricated simulation frameworks that rely on knowledge of complex ecosystem processes to simulate forest growth and succession (Taylor et al., 2008). These mechanistic or process based models are modeled at diverse spatial resolutions, as small as a leaf for photosynthesis models (Landsberg, 2003), or as large as multiple forest stands (Mladenoff, 2004). Process based models, particularly in the fields of carbon cycling (Larocque et al., 2008; Mäkelä et al., 2000) and forest disturbance (Seidl et al., 2011) have proven successful, and led to higher confidence in landscape level simulations that are able to integrate climate change into their predictions

(Duveneck et al., 2014). Mechanistic models though suffer from complexity which limits the extent and scale that can be modeled due to computational demand, as well as the availability of numerous difficult to measure inputs (Taylor et al., 2009; Weiskittel et al.,

2011).

For studies of mixed species forest types over a large study area, bioclimatic envelopes are a more suitable tactic to understand ecosystem climatic relationships, if reliable

12 empirical data is available. The focus of this climate study is not the process by which we arrive at a future landscape, but rather what the landscape might look like under different climate scenarios, obviating the necessity of a mechanistic model (Taylor et al., 2009). The spruce-fir forest, expressed as different community types across the Acadian landscape, would be difficult to capture in a mechanistic model at this scale. Undoubtedly, the abundant additional hardwoods and softwoods species that compose and interact with the spruce-fir forest types would be difficult to parameterize, and computational ability to initialize and predict a study area of 23,750,190 ha is unavailable. While bioclimatic envelopes do not account for disturbance, competition, and other filter factors determining a species presence on the landscape, it is a reliable first step in identifying a broader range of current and future suitable habitats (Heikkinen et al., 2006). Additionally, the comparison and integration of bioclimatic envelopes with process based models is able to elucidate model differences as well as ecosystem processes, while coming to a consensus on predictive futures (Kearney and Porter, 2009; Keith et al., 2008).

1.5. The Dependent Variable

The decision of which dependent variable to use in species distribution modeling is based upon the desired product and management implications of the research. A quick literature review reveals that a binary variable of presence or absence is the most commonly used, and thus, ongoing research benefits from vast information about the successes and failures of these models (Araújo et al., 2005; Elith et al., 2010; Graham et al.,

2007; Guisan et al., 2007; Heikkinen et al., 2006; Segurado and Araujo, 2004). Measurement of abundance have gained considerable popularity though, particularly in the world of forestry (Iverson et al., 2008; Prasad et al., 2006) and other plant species models (Kent and

Coker, 1992). Amongst these abundance variables, arguments persist about the practicality of different measurements, and the trade-off between model efficiency and accuracy.

The reasons for the popularity of the presence/absence dependent variable are simple. Mainly, this is the most commonly collected piece of data, and such a direct measurement leaves little room for human error. For landscape level studies which desire to characterize a species across its entire range, often numerous organizations or researchers might contribute to the model dataset. Though considerations still need to be taken into account for different sampling protocols, such as the frequency of data collection locations (Guisan et al., 2007; Luoto et al., 2005), utilizing datasets from different organizations is much simpler with the presence/absence variable. Additionally, unique datasets, such as pollen cores used in palynology studies (Williams et al., 2013), herbarium samples (Mathews and Bonser, 2005), or witness tree surveys recorded in the U.S. at the time of European settlement (Hanberry et al., 2012; Tinner et al., 2013), where abundance data is difficult to calculate, can be used in SDMs to highlight differences in realized niches

(Kearney and Porter, 2009) and the reallocation of species’ distributions in response to past climate change. Numerous modeling techniques easily accommodate the presence/absence variable, including the Ecological Niche Factor Analysis (ENFA), CARTs, neural networks, generalized linear models (GLMs), and generalized additive models (GAMs), furthering its popularity (Segurado and Araujo, 2004). CARTs have proven the most successful at accurately linking species’ distribution with climate variables (Guisan et al., 2007; Prasad et al., 2006; Segurado and Araujo, 2004). Additionally, with presence/absence modeling, balancing the data, so that errors are concentrated in favor of falsely predicting presence when absent, as opposed to absences when present, is straightforward (Joyce and Rehfeldt,

2013). With regards to endangered ecosystems, accidentally identifying regions for conservation greatly outweighs the risk of missing potential zones for refugia (Guisan et al.,

2013).

Abundance variables have gained particularly popularity in the world of species distribution modeling for forest species (Iverson et al., 2008). This is largely due to the availability of consistently measured, uniformly distributed plot networks across the landscape, such as the FIA program in the U.S., maintained by USFS. Similar datasets exist provincially in Canada (Porter et al., 2001; Prince Edward Island Department of Agriculture and Forestry, 2002; Townsend, 2004), and vary by country throughout Europe (Guisan et al.,

2007). The origins of these datasets are rooted in the economic importance of countries’ timber supplies (Bechtold and Patterson, 2005), and thus tree species are in an unique position in regards to abundance species distribution modeling. While abundance measures are often outputs in mechanistic models, the use of a continuous predictor in statistical climate modeling was difficult until the advent of CARTs (Iverson and Prasad, 1998). While balancing a dataset with a continuous variable will still help increase model accuracy, abundance models often suffer from high errors of statistical measurement (i.e. R2) because it is difficult to pinpoint exact, but varied, values across a landscape. Despite this, these models have proven immensely useful since they have the ability to reflect the sensitivity of each species to environmental gradients at their respective range boundaries, as well as depicting the core of species’ ranges (Iverson et al., 2011).

The most frequently employed abundance variable in similar studies is the importance variable (IV), which is a combined metric of both proportional basal area (BA;

15 m2 ha-1) and stem count (TPH; trees ha-1 (TPH)), and is defined in Curtis and McIntosh

(1951). The concept of the IV is that many small trees of the same species, or a few mature trees in the upper canopy, would have a similar value per unit. In regards to the species used in this study, areas of high stem count tend to simultaneously occur in areas of high basal area (Seymour, 1992). In theory, locations with a higher predicted IV are better candidates for conservation (Iverson et al., 2010). Accuracy in regards to exact values are not as important, as long as relative patterns across the landscape are achieved, and locations for conservation can be prioritized. As an alternative to direct abundance measure, the likelihood output from presence/absence CART modeling has been suggested as computationally more efficient way to calculate and display these relative patterns (Joyce and Rehfeldt, 2013). Points with a greater probability of being selected as suitable habitat are more likely to contain the species, as there is a direct relationship between greater habitat suitability and species occurrence. This is an important interpretation of presence/absence models in that it allows these models to reflect the core distribution of the species and act as a surrogate for abundance modeling.

Both presence/absence and abundance variables seek to help land managers select the best land for conservation in the face of shifting species distributions due to climate change. Presence/absence models are easier to generate and to interpret, while abundance variables help to pinpoint locations of greater habitat suitability. Neither of these types of variables assist land managers in the active management of land, nor assist in the dynamic process of a changing landscape as the climate alters. Forestry in particular, as a sect of land management that actively manages forest for multiple objectives, including timber production, wildlife habitat, and recreational opportunities, needs guidelines and tools on

16 how to manage forests under varying conditions. Density management diagrams (DMDs), which graphically represent the relationship between average tree size and stand density in forests, have long served as an important tool in making predictions about future stand development based on size-density relationships (Jack and Long, 1996). Integral to designing

DMDs is the concept of the stand-density index (SDI; Reineke, 1933), a comparative measurement that provides the degree to which a stand is achieving full site occupancy based upon the maximum size-density relationship (SDImax) (Zeide, 2005).

Traditionally, SDImax has been estimated through the visual observation of fully stocked stands, but recent research has focused on the statistical prediction of SDImax through different modeling techniques including modified linear regression (Solomon and

Zhang, 2002), nonlinear regression (Yang and Titus, 2002), and quantile regression (Zhang et al., 2013). Not only are the SDImax and DMDs universally used forestry tools, they are also particularly key for managing for forests in the face of climate change. Density management has been suggested as the single best way to achieve healthy forests, by reducing density to decrease moisture and nutrient stress caused by competition (Chmura et al., 2011), and therefore reducing vulnerability to wildfire and disease outbreak (Noss, 2001), known agents of acute mass mortality in climate stressed ecosystems (Allen et al., 2010).

Integrating the results of a landscape level SDImax prediction into a climate model has not been attempted at the time of this study. Careful considerations need to be taken in regards to compounded risk of error associated with the stacking of model results.

1.6. Objectives

It is clear that the spruce-fir forest type of the Acadian Forest is an unique assemblage of species that provides invaluable economic and ecological resources. Land managers need accurate information in order to conserve and manage for changes to this ecosystem under different models of climate change, and different dependent variables provide different types of information. Modeling alternative dependent variables for different species though is rarely performed due to the lack of data availability, thus missing the opportunity to inspect species’ performance to different response variables and to study the different implications these modeling outcomes could have on conservation decisions. Thus, there is a need to compare these variables on the same landscape and to understand their implications, while also exploring innovative modeling techniques.

The broad objectives of research documented in this thesis were:

1. To explore new data and modeling techniques for SDMs. This includes the impact of

higher spatial resolution, and the impact of the use of an international dataset

composed from numerous current and historical sources, on predictive accuracy,

and the ability of newly developed statistical techniques to predict important

variables for forest management, such as SDImax.

2. To characterize the distribution and abundance of the important species spruce-fir

forest, while comparing the usefulness of both presence/absence and abundance

models, as well as alternatives, for conservation decisions.

3. To compare and illustrate the differences between the results and application of

directly calculated variables useful for passive management versus predicted

variables useful for the active management of forests.

CHAPTER 2

MODELING AND FORECASTING EASTERN NORTH AMERICAN SPRUCE-FIR

OCCURRENCE/ABUNDANCE UNDER CURRENT AND FUTURE CLIMATE CONDITIONS

2.1. Abstract

The spruce-fir (Picea-Abies) forest type of the Acadian Region is at risk of disappearing from the United States and parts of Canada due to climate change and associated impacts.

This valuable ecosystem provides habitat to wildlife of both local and national conservation concern, and sustains regional economies. Managing for the multiple resources provided by this ecosystem requires accurate forecasting across international boundaries in the face of expected tree species distribution shifts. This analysis linked species specific data with climate and topographic variables using the nonparametric random forest algorithm, to generate models that accurately predicted changes in species distribution under different models of climate change. Previous analyses of these species were limited due to coarse spatial and temporal resolution of analyses, the dependent variable employed, and geopolitical limitations associated with fully characterizing the species’ ranges, particularly into Canada. A database consisting of over 10 million individual field observations of tree occurrence and abundance (defined as basal area, stem density, and importance value) was compiled from the species’ current and potential range. When compared to other approaches, the occurrence models were able to accurately determine current distribution.

Area under receiver operator curve (AUC) values for models averaged 0.99 ± 0.01 (mean ±

SD), well above the accepted standard for excellent model performance. Abundance modeling results varied, with model performance contingent upon individual species’ characteristics. Black spruce (Picea mariana (Miller) B.S.P.) responded the best to

19 abundance modeling, while red spruce (Picea rubens Sarg.) and white spruce (Picea glauca

(Moench) Voss) distribution were most accurately estimated through presence/absence models. The addition of historical tree data revealed supplementary suitable habitat along the southern edge of species’ ranges, due to marginal dynamics potentially overlooked by approaches relying solely on current inventories. Future predictions suggest an almost complete extirpation of suitable spruce-fir habitat from the United States by the year 2090, with the exception of locations at high altitudes in the Adirondacks and along the

Appalachian Mountain chain in New Hampshire and Maine. Areas of large future suitable habitat are predicted for interior and peninsular Newfoundland and along the Gulf of St.

Lawrence in Québec, including the northeastern tip of the Gaspé Peninsula, the Côte-Nord region, and Anticosti Island. These outcomes will help public and private land managers evaluate multiple alternative scenarios in which ecosystem perseverance, economic profitability, and concerns for wildlife habitat can be accounted for in the face of uncertainty.

2.2. Introduction

According to the latest report by the Intergovernmental Panel on Climate Change

(IPCC), global surface temperatures are likely to rise between 0.3 and 4.8°C by the end of the 21st century (Stocker et al., 2013). Additionally, the last three decades are likely the warmest 30-year period of the previous 1400 years, with a temperature increase of 0.7°C in that time. This increase in temperatures has cascading effects on sea surface temperatures, annual precipitation, glacier and ice sheet volume, and many more aspects of the global climate system. These changes to climate are unsurprisingly reflected in species’ distributions and ecosystems’ configurations. It is recognized that as temperatures rise

20 species’ geographic distributions generally shift poleward and upward in altitude (Harsch et al., 2009; Lenoir et al., 2008; Parmesan, 2006). Paleoecological evidence confirms that temperature shifts as little as 1°C led to significant forest reconfigurations as little as 1,000 years ago (Lindbladh et al., 2003; Schauffler and Jacobson, 2002). Currently, transformations are already being witnessed, with one meta-analysis of mobile organisms estimating a median latitudinal migration of 16.9 km per decade and a median shift to higher elevations of 11 m per decade (Chen et al., 2011). Climate impacts on sessile flora, such as forests, are still being evaluated, as response to climate change is complex, relying on the interactive effects of both temperature and precipitation changes (Parmesan, 2006). Numerous studies have documented the shift of forest habitat (Kelly and Goulden, 2008; Lenoir et al., 2008) upward in altitude, or the loss of ecosystems altogether (Condit et al., 1996), due to climate change. Rapid migration potential is limited, and shifts in the suitability of habitat conditions

(Iverson et al., 2008), or the reconfiguration of forest structure, composition, and productivity (Dolanc et al., 2013; Mohan et al., 2009), are a more immediate common outcome of climate warming.

The Acadian Region of North America is expected to have hotter summers and shorter winters marked by more rain and less snow (Jacobson et al., 2009). Projected future changes are consistent with a warmer climate, including shrinking snow cover, more frequent droughts, and extended periods of low hydrological flows in the summer (Hayhoe et al., 2007). Summertime precipitation is projected to decrease on the Acadian coastline and inland, but increase along the Canadian border (Anderson et al., 2010; Hayhoe et al.,

2008). Meanwhile, evaporation is expected to increase in most of the region, resulting in lower soil moisture content and higher humidity (Anderson et al., 2010). Already, overall

21 average temperatures have increased by 0.37 to 0.43°C per decade since 1965, with greater temperature increases in the winter, and the amount of days with snow on the ground has decreased by up to 25 days (Huntington et al., 2009; Wake et al., 2006). This change in climate is already being manifested in the regional redistribution of forests, with one study reporting an upward shift of 91 to 119 m in the montane northern hardwood-boreal forest ecotone in Vermont (Beckage et al., 2008).

Several other coarse scale analyses have addressed the potential reduction or loss of species richness in Northeastern United States (U.S.) as species and communities migrate northward (Hansen et al., 2001; Iverson et al., 2008; Tang and Beckage, 2010). Of particular concern within the Acadian Forest, is the spruce-fir (Picea-Abies) forest type, as the primary tree species in this forest, red spruce (Picea rubens Sarg.), black spruce (Picea mariana

(Miller) B.S.P), white spruce (Picea glauca (Moench) Voss), and balsam fir (Abies balsamea

L.), prefer cooler and moister conditions associated with northern latitudes and sensitive high alpine and coastal areas. Previous climate models have predicted range contraction of up to 400 km north (Iverson et al., 2008) and a possible reduction of 97-100% of suitable spruce-fir habitat in the U.S. in the next 100 years (Hansen et al., 2001). Refugia locations in

New England are predicted to be restricted to high elevations or inland along the United

States-Canada border (Tang and Beckage, 2010). The risk of this shrinking habitat is further compounded by the fact that several species of local (e.g. spruce grouse (Dendragapus canadensis canace)) and national concern (e.g., Bicknell’s Thrush (Catharus bicknelli),

Canadian Lynx (Lynx canadensis)) rely on the spruce-fir forest and that this habitat is already considered uniformly rare in Maine and endangered in New York (Noss et al., 1994). These previous studies that have predicted range contraction of the spruce-fir forest type have

22 been limited by the absence of data that fully characterizes the species’ relationships with the environment in the northern portion of their range, as it reaches across international boundaries, preventing range wide modeling and monitoring. The absence of this data not only limits understanding of species and climate associations, but also prohibits recognizing future suitable habitat for forest communities and their associated wildlife.

For studies of mixed species forest types over a large study area, developing statistical models that link individual species’ distributions with important environmental variables, has been suggested as an appropriate tactic towards understanding ecosystem climatic relationships, if reliable empirical data is available. Also known as species distribution models (SDMs), ecological niche models, and bioclimatic envelopes, this method is an empirical based approach to correlating the presence of species to climatic variables, assuming the hypothesis that the best indicator of a species realized niche is its current distribution (Pearson and Dawson, 2003). While bioclimatic envelopes do not account for disturbance, competition, and other filter factors determining a species presence on the landscape, it is a reliable first step in identifying a broader range of current and future suitable habitats (Heikkinen et al., 2006). Accurate and comprehensive datasets are necessary in order to fully characterize species relationships with climate. Empirical data utilized in species distribution modeling typically rely on a single data source to describe species relationships with their environment. Sources range from records obtained from

Herbaria, Museums or Atlases (Austin, 2007; Graham et al., 2007) to systematic national inventories of trees (Guisan et al., 2007; Iverson et al., 2008) and other species (e.g.. butterflies (Luoto et al., 2005)). Within the U.S., analyses of tree species have typically relied on the Forest Inventory and Analysis (FIA) program maintained by the U.S. Forest Service

(USFS) (Iverson et al., 2008). While it appears that this dataset can accurately delineate the presence of tree species at a coarse resolution, particularly those with large uniform distributions, the ability of this dataset to precisely capture species with a small specific niche or low abundance, in a mixed species landscape, at a fine resolution, is unsure.

Additional, obvious limitations arise from the ability of a single national inventory to correctly describe species’ ranges that cross international boundaries.

Furthermore, in areas of continual intense anthropogenic disturbance and settlement, where forest habitats have been altered or excised, the temporal range of FIA and other datasets do not include data prior to disturbance. Known distribution is usually limited to information collected after 1900, and primarily after 1950 (Elith et al., 2006). While the FIA was established in 1930, it did not begin regular inventory until 1998 (Bechtold and

Patterson, 2005). For North American tree species, historical records collected at the time of

European settlement are widely available. Previous analyses based on historical tree data have shown changes between historical and current forest species composition and abundance, resulting from logging, fire suppression, and other anthropogenic disturbances

(Cogbill et al., 2002; Hanberry et al., 2012). Less studied is how land-use has affected suitable habitat for individual tree species, or what effect this might have on species’ bioclimatic envelopes and subsequent assessments of climate impacts on future suitable habitat (Tinner et al., 2013).

Species’ distributions have primarily been defined in previous studies through presence/absence data (Elith et al., 2010; Guisan et al., 2007), as this type of data is widely available. Abundance variables (i.e. basal area (BA), stem density, importance value (IV)) have gained particularly popularity in the world of species distribution modeling for forest

24 species (Iverson et al., 2008), due to the availability of consistently measured, uniformly distributed plot networks across the landscape, such as the FIA program. Modeling for these different types of dependent variables serves slightly different purposes. Presence/absence models benefit from user-generated balancing that can concentrate error in favor of falsely predicting presences when absent as opposed to absences when present (Joyce and

Rehfeldt, 2013). With regards to endangered ecosystems, identifying regions with current conservation value greatly outweighs the risk of eliminating potential zones for refugia

(Guisan et al., 2013). Abundance variables are seen as useful because they have the ability to reflect the sensitivity of each species to environmental gradients at their respective range boundaries, as well as depicting the core of species’ ranges (Iverson et al., 2011). In theory, locations with a higher predicted abundance are better candidates for conservation (Iverson et al., 2010). Accuracy in regards to exact values are not as important, as long as relative patterns across the landscape are achieved, and locations for conservation can be prioritized. Regardless of modeling intent, inherent species characteristics, such as prevalence and range size, are thought to chiefly influence model performance (Guisan et al., 2007; Luoto et al., 2005; Segurado and Araujo, 2004), and abundance versus presence/absence data might outperform one another under different circumstances.

As an alternative to direct abundance measures, the likelihood output from presence/absence models has been suggested as computationally more efficient way to calculate and display these relative patterns (Joyce and Rehfeldt, 2013). It is inferred that points with a greater probability of being selected as suitable habitat are more likely to contain the species, as there is a direct relationship between greater habitat suitability and species occurrence. This is an important interpretation of presence/absence models in that

25 it allows these models to reflect the core distribution of the species and act as a surrogate for abundance modeling. Modeling alternative dependent variables for different species though is rarely performed due to the lack of data availability, thus missing the opportunity to inspect species’ performance to different response variables and to study the different implications these modeling outcomes could have on conservation decisions.

In order to capture the full range of spruce-fir species’ relationships with their environment across the entire Acadian Region, individual bioclimatic envelopes were developed with a comprehensive dataset including resources from both the U.S. and

Canada. Both presence/absence, likelihood, and abundance variables (i.e. relative BA, relative stem density, IV) were examined to evaluate the ability of bioclimatic envelopes to accurately model species’ distributions and to reflect cores of distribution. These models were constructed with and without historical observations to observe the effect that obfuscated habitat ranges might have on species’ bioclimatic profiles. Models were built at a fine resolution (1 km²) in order to assist in identifying areas of potential refugia at the extremes of species’ habitats under different models of climate change. This fine resolution will be of more practical use to land managers than previous coarse-resolution models. The specific objectives of this study were to: (1) develop species-specific current distribution models using contemporary data; (2) compare predictions when contemporary and/or historical data are used; (3) evaluate alternative methods for estimating the current distribution; and (4) generate predictive maps of future distribution.

2.3. Methods

2.3.1. Study Area

The species considered for this study included red spruce, black spruce, white spruce, and balsam fir. In order to fully understand these species’ relationship with their environment, the study area extended beyond the boundaries of the Acadian Forest to include the southern extent of species’ ranges (Figure 2.1). As species migrate northward it is expected they will exhibit similar associations with their environment in the Acadian

Region as they do today to the south. For example, isolated red spruce populations, in conjunction with Fraser fir (Abies fraseri Pursh) Poir), are located throughout the southern

Appalachians at elevations above 1400-1600 m (Stephenson and Adams, 1984). These populations are thought to be relics from the last ice age, with shrinking suitable habitat as climate warms (Oosting and Billings, 1951). The breadth of the study area included data from ecoregions north of the Acadian Forest as well, but the northern edge of multiple

Figure 2.1. Study area overlaid with World Wildlife Fund (WWF) ecoregions.

27 species’ ranges were not included in this analysis. Black spruce, white spruce, and balsam fir ranges extend well into the Canadian taiga, a region where little tree data has been collected. Classifying this edge was not thought to have an impact on describing species’ distributions for the Acadian Region under current or future climate scenarios, particularly given the elevational equivalents for these bioclimatic conditions contained within the mountainous portions of the study area.

The New England-Acadian Forest terrestrial ecoregion defined by the World Wildlife

Fund (WWF) was used to cartographically delineate this region (Olson et al., 2001). Data overlapped with 17 additional ecoregions in this analysis. The Allegheny highlands,

Appalachian Blue Ridge forests, Appalachian mixed mesophytic forests, Central Canadian

Shield, Eastern Canadian forests, Eastern forest-boreal transition, Eastern Great Lakes lowland forests, Gulf of St. Lawrence lowland forests, Newfoundland highland forests,

Northeastern coastal forests, and Southern Hudson Bay taiga coincide with at least one of the species’ ranges. Data from six other regions including the Atlantic coastal pine barrens,

Central U.S. hardwood forests, Middle Atlantic coastal forests, South Avalon-Burin oceanic barrens, Southern Great Lakes forests, and Southeastern mixed forests was used to supply information about climate characteristics outside of the species’ ranges.

2.3.2. Tree Data

Observations, including individual tree species and diameter at breast height (dbh), were gathered from various agencies in the U.S .and Canada to provide detailed coverage of the study area. Strict attention was paid to sampling protocols used by each organization in order to consistently calculate the necessary variables for analyses. Details about the protocols used by each organization are in Appendix A. Four dependent variables were

28 determined from this data. These were measures of presence or absence for each species and three measures of relative abundance, including stem density (trees ha-1), BA (m2 ha-1), and an IV. The IV is a combined metric of proportional stem density and BA defined in Curtis and McIntosh (1951). All variables were calculated at the plot level and expanded to one ha.

A threshold of 10 cm and greater for individual tree dbh was used to calculate these values.

This threshold was used to target the core of distribution. Preliminary analysis performed using smaller dbh thresholds indicated only small changes in predictions of suitable habitat

(Appendix B). The primary focus was to collect data sampled from spruce-fir habitat, but observations from different forest types within or near to the Acadian Forest were also obtained. This absence data was used to train models to distinguish whether spruce-fir will be present or not, particularly in areas with similar climatic and geographic profiles, but different forest types.

2.3.2.1. Contemporary Tree Data

10,493,619 observations on 248,821 plots were collected to provide details about the contemporary distribution of species. The data collection period spanned from 1955 to

2012, with the majority collected after 1980 (85%). The New Brunswick Department of

Natural Resources, the Newfoundland Forest Service, the Nova Scotia Department of

Natural Resource, the Québec Ministry of Natural Resource, and the Prince Edward Island

Department of Agriculture and Forestry provided coverage of Canada. In the United States, the Cooperative Forestry Research Unit, the Massachusetts Department of Conservation and Recreation, the National Park Service, the New Hampshire Division of Parks and Lands, the USFS, the University of Maine, the University of Massachusetts, the Vermont Center for

Ecostudies, and the Vermont Monitoring Cooperative provided data. Data from the USFS

FIA was primarily utilized to provide wider coverage of absence data. The USFS FIA provided a substantial amount of the data in this analysis and predictions generated using solely FIA data are presented in Appendix C as a direct comparison to previous similar analyses

(Iverson et al., 2008). In short, the use of only FIA data produced similar model fit statistics, but accurate predictions were not obtained across the full range of the species, particularly in Canada (Appendix C).

2.3.2.2. Historical Tree Data

1,342 historical tree observations from 778 plots were obtained from a database maintained by Charles Cogbill (Cogbill, 2000). This data was originally collected between

1623 and 1869 and represents tree composition at the time of European settlement in the

New England states and New York. This land was surveyed at the time of division into 40 –

60 ha lots by proprietors, with the largest tree at the corner of each lot recorded as a demarcation boundary (Cogbill, 2000). Though sampling methods were often poorly documented, these observations are thought to be representative of town wide composition at the time of collection, as they were collected on a grid pattern (Cogbill et al.,

2002). Only presence/absence, not abundance, can be garnered from this data. Inclusion of this data provided a unique opportunity to account for habitats and regions that may have historically supported spruce-fir species prior to extirpation by land use or other factors.

2.3.3. Climate Data

Climate data was collected from Moscow Forest Science Laboratory climate database available online at http://forest.moscowfsl.wsu.edu/climate/ (download date 05

January 2014). Climate data was derived by applying thin-plate smoothing spline procedures that extrapolate data from discrete weather stations to specific plot points with

30 corresponding elevation (Rehfeldt, 2006). Current climate data was normalized for a thirty year period (1960-1990) and was based on weather station data for about 15,000 locations for precipitation and 12,000 for temperature (Joyce and Rehfeldt, 2013). 33 climatic variables were used in analysis, which have been shown to be effective in previous analyses

(e.g. Joyce and Rehfeldt, 2013) (Table 2.1). Sixteen of these variables are direct measurement of climate, while the remaining seventeen are second-order interactions.

2.3.4. Topographic Data

Topographic variables were used to model species occurrence and abundance in order to capture discrete landscape features that influence species’ dynamics and life history outcomes, and also to capture effects that terrain features might have on microclimate. Elevation, slope, and aspect data were collected, if available, from the original data source. If not available, elevation data was extracted from the 30 m resolution national elevation dataset (NED) generated by the United States Geological Service (USGS) available at http://viewer.nationalmap.gov/viewer/ (download date 12 February 2013) and from the

30 m resolution digital elevation dataset made available through the Canadian Council on

Geomatics (CCOG) available at http://www.geobase.ca/geobase/en/find.do?produit=cded

(download date 3 March 2014). Slope and aspect were derived from the NED using the raster package (Hijmans, 2014) available through R statistical software (R Core Team, 2013).

Table 2.1. Description of climate variables used in analysis. Mean, standard deviation (S.D.), minimum (Min), and maximum (Max) values are listed for both the plots used in this analysis and the entire study area. Climate variables in bold represent those which were used to construct the absence sampling hypervolume. Plots Study Area Acronym Definition Mean S.D. Min Max Mean S.D. Min Max Julian date of when the D100 number of days 84.9 39.3 17.0 188.0 114.8 40.6 17.0 197.0 above 5°C reaches 100 Annual number of days below DD0 0°C based on 455.6 666.7 0.0 3233.0 975.8 898.7 0.0 3480.0 mean monthly temperature Annual days above 5°C based DD5 on mean 3127.8 1208.4 503.0 5358.0 2267.2 1222.4 356.0 5372.0 monthly temperature Julian date of first freezing FDAY 288.9 21.4 238.0 348.0 274.1 22.2 237.0 349.0 temperature in autumn Frost free period FFP 174.7 48.0 59.0 298.0 141.9 49.0 58.0 298.0 length Mean number of days above GSDD5 2630.6 1072.8 311.0 4852.0 1883.5 1084.3 240.0 4858.0 5°C between SDAY and FDAY Growing season (April - GSP 627.3 72.8 353.0 1108.0 588.5 70.0 323.0 1128.0 September) precipitation Mean annual MAP 1203.5 180.7 656.0 2217.0 1110.9 188.9 654.0 2374.0 precipitation Mean annual MAT 11.4 5.7 -5.2 19.7 7.0 6.4 -6.4 19.7 temperature Mean maximum temperature in MMAX 28.9 3.9 14.1 33.9 25.9 4.8 12.4 33.9 the warmest month Mean minimum temperature in MMIN -7.3 7.8 -32.2 4.6 -13.0 9.0 -32.2 4.7 the coldest month Annual number of days below 0°C based on MINDD0 942.6 958.0 20.0 4696.0 1685.2 1226.1 19.0 4943.0 mean minimum monthly temperature

Table 2.1. continued Plots Study Area Acronym Definition Mean S.D. Min Max Mean S.D. Min Max Mean temperature in MTCM -1.5 7.9 -24.6 10.7 -7.4 8.9 -25.5 10.8 the coldest month Mean temperature in MTWM 22.8 3.9 10.4 27.6 20.0 4.5 8.9 27.6 the warmest month Julian date of last freezing SDAY 113.2 25.8 52.0 184.0 131.5 26.9 52.0 187.0 temperature in spring TDIFF MTWM-MTCM 24.3 4.4 16.7 37.1 27.3 5.0 16.6 37.2 Interactions Annual dryess ADI index: 0.05 0.01 0.01 0.06 0.04 0.01 0.01 0.06 (DD5)0.5/MAP Annual dryness ADIMINDD0 & cold index: 37.4 32.8 1.1 176.4 61.9 40.7 1.1 182.4 ADI * MINDD0 (DD5 * DD5MTCM 4.7 19.1 -26.0 57.3 -6.5 15.5 -26.5 58.0 MTCM)/1000 (GSP * GSPDD5 2002.6 885.0 240.4 4216.9 1383.1 853.5 198.8 4234.4 DD5)/1000 (GSP * GSPMTCM -0.6 4.8 -15.9 8.4 -4.0 5.0 -17.1 8.5 MTCM)/1000 (GSP * GSPTD 150.6 22.4 88.3 282.7 158.8 24.5 80.8 295.7 TDIFF)/100 (MAP * MAPDD5 3877.5 1787.8 531.2 8236.7 2649.3 1731.4 427.7 8289.3 DD5)/1000 (MAP * MAPMTCM -0.9 8.9 -28.9 14.3 -7.0 9.0 -36.1 14.4 MTCM)/1000 (MAP * MAPTD 287.2 39.3 190.0 508.5 296.4 36.2 179.5 593.5 TDIFF)/100 MTCMGSP MTCM/GSP 0.0 0.0 -0.1 0.0 0.0 0.0 -0.1 0.0 MTCMMAP MTCM/MAP 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 PRATIO GSP/MAP 0.5 0.0 0.4 0.7 0.5 0.0 0.4 0.7 PRDD5 PRATIO * DD5 1626.3 616.6 227.7 3259.0 1196.0 618.1 166.0 3267.0 PRMTCM PRATIO * MTCM -0.9 4.4 -15.8 6.5 -4.2 5.1 -16.1 6.6 Summer dryness SDI index: 0.1 0.0 0.0 0.1 0.1 0.0 0.0 0.1 (GSDD5)0.5/GSP TDGSP TDIFF/GSP 0.0 0.0 0.0 0.1 0.0 0.0 0.0 0.1 TDMAP TDIFF/MAP 0.0 0.0 0.0 0.1 0.0 0.0 0.0 0.1

33 curvature. These variables were assumed to capture effects not reflected in the climate variables such as soil drainage, exposure, and solar radiation profiles.

2.3.5. Species-Specific Distribution Model Development

Four different dependent variables were used to construct species’ bioclimatic profiles. Species-specific presence/absence models were constructed with and without historical tree data to evaluate differences with the inclusion of this data. All models were constructed using the random forest algorithm (Liaw and Wiener, 2002) available in R (R

Core Team, 2013). Random forest can create classification or regression trees. Classification trees have been shown to have high predictive accuracy for presence/absence species distribution modeling (Elith et al., 2010; Guisan et al., 2007), while the regression component of random forest has been used in abundance modeling (Iverson et al., 2008).

The classification and regression components of random forest are very similar, but differences lie in how many random independent variables are selected at each node (i.e. square root of all independent variables for classification, one-third for regression) and the default node size at each split (i.e. one for classification, five for regression).

Random forest is an ensemble learning model that aggregates the results of multiple unique trees. Each tree is generated by sub-sampling two-thirds of the complete dataset and then recursively partitioning the data by choosing the optimal predictor variable for splitting the data at each node (Liaw and Wiener, 2002). Random forest is unique in that at each node a subset of the independent variables is selected. This added layer of randomness reduces correlation between trees and thus decreases total forest error rate

(Breiman, 2001). Additionally, selecting from a subset of independent variables increases computational efficiency making this algorithm ideal for large datasets with a multi-

34 dimensional independent variable space. Partitioning is complete once error can no longer be reduced and multiple terminal nodes are reached. The result is a tree that predicts for the dependent variable at each terminal node, by means of deriving the average response value (regression) or most common response (classification) from the observations within this node using a piecewise constant prediction function (Joyce and Rehfeldt, 2013; Strobl et al., 2009). Data points are then predicted by aggregating the votes from each tree. For classification, the majority of votes determines class output and for regression an average value is calculated.

Prevalence, or the percent of individuals present in the dataset, for each species was relatively low across the represented landscape. In order to address the concern that the random forest algorithm relies on equal representation across classes for accurate prediction (Joyce and Rehfeldt, 2013), absence data was down sampled to represent approximately 50% of the dataset in the presence/absence models and 20% in the abundance models of species (Chen et al., 2004). Furthermore, the number of present or abundance data points were duplicated prior to absence sampling and analysis. Increasing prevalence within a dataset relative to the actual incidence across the landscape decreases erroneous predictions of absence without violating any basic statistical assumptions

(Pearson and Dawson, 2003). This was considered important for the study, in which the goal was to identify future suitable habitats of an at-risk ecosystem.

It is important to provide random forest algorithms with absence data in order to train the model to distinguish not only the limits of the species’ range boundaries, but also to differentiate between areas with similar abiotic features but with dissimilar species abundance or composition. The construction of each model dataset varied, but preliminary

35 analyses showed an approximate ratio of 50-50, presence to absence, and 80-20 for regression models, provided the most accurate results. Approximately half of the absence data were sampled from areas determined to be climatically similar to the presence or abundance data for each species. To establish climatic similarity, an eighteen variable hypervolume was defined per species and expanded by 0.01 standard deviation in all directions (Table 2.1, in bold) (Joyce and Rehfeldt, 2013). To complete the dataset, additional “outside” absence data were sampled from beyond the established hypervolume.

Absence data were either randomly sampled without replacement or randomly sampled within strata defined by ecoregion, depending upon model performance.

For each dataset, a random forest consisting of 500 trees was ran five times. The most important variables were determined using the unscaled permutation accuracy importance measure based on the VarImp function option in the random forest package.

This measure is a calculation of the mean decrease in accuracy for classification, or the mean decrease in node impurity for regression, when a variable in the tree is randomly permutated to another variable. Permuted variables that result in a higher decrease in precision are considered more important. The unscaled computation of this measure was used because the scaled measure has shown preference of correlated predicted variables

(Strobl et al., 2007) and results provide greater predictive accuracy. Preliminary analyses showed that reducing the complete array of 43 variables to the top eight, five, and two predictors, resulted in an average 14.6%, 16.2%, and 47.3% increase in out of bag (OOB) error, respectively, for classification models and an average 4.9%, 5.8%, and 9.1% decrease in R2, respectively, for regression models. The five most important variables were selected for each model as this number was considered a parsimonious balance between model

36 accuracy, computational efficiency, and the ability to describe each species’ relationship with its environment. Final models were generated using the most important variables in a random forest with 500 trees.

2.3.6. Model Evaluation and Comparison

Measures of accuracy considered in this study for presence/absence models were area under receiver operator curve (AUC) and OOB error. A pseudo R2 was used for regression models. All models were predicted across the current landscape. Kappa values were used to compare predicted current distribution against actual distribution using the

Map Comparison Kit (Visser and de Nijs, 2006).

The random forest algorithm reserves one-third of the model dataset, referred to as the OOB sample, for each tree that is constructed. This sample is used to internally estimate the precision of the tree constructed by running the sample down the tree and recording the accuracy of each data point’s value (Breiman, 2001). For regression models, the mean square error (MSE) as well as a “pseudo R2” is calculated and reported to determine accuracy. Random forest’s R2 differs from the traditional R2 in that the variance is calculated by dividing by n, as opposed to n-1.

For classification models, the OOB error is calculated as a proportion of misclassifications per data point relative to the total number of observations in the forest.

The OOB error is represented as a confusion matrix, in which two types of misclassifications can be calculated, errors of commission and errors of omission. Errors of commission refer to erroneous predictions of presence. Models with a high commission error are referred to as having low specificity. Errors of omission refer to erroneous prediction of absence.

Models with a high omission error are referred to as having low sensitivity (Pearson et al.,

2004). As mentioned before, having a greater prevalence will increase overprediction and lower omission rates. These two metrics are calculated independently of one another and can be misleading as to overall model performance (Pearson and Dawson, 2003). Thus, OOB classification error rates were used as an index to determine the best prevalence rate and sampling scheme for each specific dataset, but not as a metric for comparing one model to another.

AUC and Cohen’s kappa statistic of similarity (kappa) are both measures that evaluate overall model agreement between predictions and observation. Kappa is a measure that corrects for agreement expected to occur by chance (Cohen, 1960), but suffers from the necessity of a user defined threshold above which the model outputs are considered present (Pearson et al., 2004). AUC assesses the full range of threshold values by plotting the tradeoff between sensitivity and specificity for any given model (Fielding and

Bell, 1997). These measures have been shown to be highly correlated to one another

(Graham et al., 2007; Pearson and Dawson, 2003) and both measures have been widely used in species modeling, although AUC has widely replaced kappa in recent years. AUC values can range from 0 to 1, with values greater than 0.5 representing model performance greater than chance (Fielding and Bell, 1997). Similar studies have considered AUC values below 0.7 as poor, between 0.7 and 0.9 as useful, and over 0.9 as good to excellent (Guisan et al., 2007).

Kappa was used in the study to compare model generated maps to actual maps of species distribution and abundance. Kappa values can range from -1 to 1, with 1 presenting perfect agreement in the distribution of categories between two maps. Values of kappa greater than 0.75 are regarded as very good, values between 0.4 and 0.75 indicate fair

38 agreement, and values less than 0.4 signify a poor relationship (Monserud and Leemans,

1992). To analyze using this method, the abundance predictions were delineated into eight equal categories representing the range of values, post-regression. Two elements of kappa further describe the locational (Kloc) and quantitative (Khist) similarities between two map objects. Kloc describes the accuracy of spatial allocation of categories by comparing the actual to expected rate relative to the maximum success rate that could be obtained if the locations of the categories in one map were rearranged. Khist is a similarity index that compares the histograms of the two maps (Pontius, 2000; Prasad et al., 2006).

Lastly, the probability prediction object for each presence/absence model was generated to examine the likelihood of occurrence, which measures the proportion of trees in the random forest object that produced a positive vote at each pixel. For example, if a point received a positive vote 400 times in a forest consisting of 500 trees, it has an 80% likelihood of occurring at that point. These probability objects have been proposed as a surrogate for abundance models in regards to their ability to reflect the core and outer limits of distribution. A default threshold of 50% for non-occurrence versus occurrence was used in this analysis, as mapping this threshold has shown strong similarity to Little’s (1971) range maps (Joyce and Rehfeldt, 2013). Probability predictions were mapped and displayed in the strata presented in Joyce and Rehfeldt (2013). Locations with a likelihood occurrence between 50 and 85% are shown in yellow and those greater than 85% are indicated in green. For comparison, actual abundance values were divided and displayed by those that are in the top 15th percentile of predicted values and those between the 50th and 85th percentile. Spearman’s rank correlation coefficient ( ) was reported to detect trends between the two predicted datasets. The Spearman rank is a nonparametric technique that

39 divides the data into ranks in order to inspect the relationship between two variables and values can range from -1 to 1 (Chok, 2010). This metric was used instead of the more popular Pearson’s correlation coefficient as inferences from do not rely on assumptions of normal distribution.

2.3.7. Predictive Mapping

Mapped predictions of future distribution for spruce and fir forest types of the

Acadian Region were generated using the output of the random forest predicted over different climate landscapes in the years 2030, 2060, and 2090. Mapping was based on

0.00833° (~1 km2) grid and generated with the raster package (Hijmans, 2014) in R. Future landscapes were acquired for each important variable through the Moscow Forest Science

Laboratory’s climate database. The ENSEMBLE representative concentration pathways 6

(RCP6) scenario, generated in affiliation with the IPCC was used to forecast future suitable habitat. Different RPCs were created by analyzing varying predicted rates of radiative forcing, as well as greenhouse gases emission rates and concentrations by the year 2100

(Stocker et al., 2013). RCP6 is a moderate scenario, the 6 referring to the radiative forcing in

2100 measured in watts m2.

2.4. Results

2.4.1. Data Characteristics

Balsam fir, white spruce, black spruce, and red spruce were located in 15.4%, 6.6%,

9.1%, and 4.1% of plots, respectively. The majority of these plots were located in the New

England-Acadian Forests, the Eastern Canadian Forests, and the Eastern forest-boreal transition (Figure 2.2). Absence data was represented across all ecoregions and accounted for 79.5% of observations and 64.9% of plots in the total dataset. A majority of the absence

40 data was provided by the USFS FIA (95.5%), while spruce-fir data was collected primarily from non-FIA sources (97.3%). The distribution of relative BA, relative stem density, and the

IV of white spruce, red spruce, and balsam fir all exhibited descending monotonic type shapes (Figure 2.3). Black spruce exhibited a relatively even distribution with higher concentrations located near zero and near one. Mean values varied, but relative abundances were higher overall for balsam fir and black spruce, and lower for white and red spruce (Table 2.2).

Table 2.1 exhibits the climatic variation of the dataset compared to the entire study area. In general, plot climate data reached the minimum and maximum values of the study area. The means of the plots were within one standard deviation of the means of study area. Plot were higher or lower than study area means due to the concentration of the collection of the data in the central portion of the study area.

2.4.2. Model Performance

Sensitivity ranged from 98.84% (balsam fir) to 99.49% (black spruce) and specificity ranged from 91.01% (black spruce) to 95.17 (red spruce) (Table 2.3). Black spruce exhibited the largest difference between these two values, while red spruce demonstrated the smallest. All AUC values were 0.99, with the exception of white spruce, which was 0.98. AUC values were well above 0.90, signifying that these models were excellent representations of their datasets.

For the abundance metrics, BA models performed slightly better than stem density or IV models (Table 2.3). Similar patterns were exhibited for each species between the abundance models. White spruce consistently displayed the lowest R2 (± MSE) (BA: 67.7 ±

0.01; Stem Density: 65.0 ± 0.02; IV: 65.3 ± 261.4) and the R2 for black spruce (BA: 87.8 ±

0.02; Stem Density: 87.5 ± 0.02; IV: 88.1 ± 352.6) was the highest. More accurate results were exhibited for the balsam fir BA and IV model (BA: 78.5 ± 0.02; IV: 76.1 ± 357.4) as opposed to the stem density model (72.0 ± 0.03). Red spruce demonstrated midrange values for all dependent variables (BA: 74.7 ± 0.01; Stem Density: 73.3 ± 0.03; IV: 73.6 ± 254.5).

The average percent difference between the actual and predicted means were 39.5% for

BA, 43.6% for stem density, and 40.5% for IV. Density plots for the observed data are overlaid with the prediction object density distribution in Figure 2.3. Overall, random forest abundance models were better at detecting mid-range values, but overestimated low abundance and underestimated high abundance on the landscape, driving down predicted mean values. Overprediction of very low values was particularly a problem with the stem density models(TPH) and the red spruce BA and IV models. The white spruce models were an exception to this, as this species occurs frequently at low abundances across the landscape. Red spruce’s actual means were the most disparate from their predicted means, while black spruces’ were the closest. Predicted midrange values for balsam fir, white spruce, and black spruce were all artificially elevated, while red spruce’s predictions consistently underpredicted values greater than zero.

Models exhibited similarity in regards to variable importance selection. The same five variables were selected for each presence/absence model, though rank varied (Table

2.3). For the abundance models, a total of seven different variables were selected, including the five that were the closest. Predicted midrange values for balsam fir, white spruce, and black spruce were all artificially elevated, while red spruce’s predictions consistently underpredicted values greater than zero.

Figure 2.2. Frequency of types of data sorted by ecoregion. Ecoregion designated by World Wildlife Fund (WWF). FIA = Forest Inventory and Analysis.

Table 2.2. Statistics of abundance values by species. Mean and standard deviation (S.D.) are included.

Relative Stem Importance Value Species Relative Basal Area Density Mean S.D. Mean S.D. Mean S.D. Balsam Fir 0.31 0.27 0.39 0.30 33.84 27.09 White Spruce 0.17 0.21 0.17 0.23 16.65 20.39 Black Spruce 0.51 0.37 0.52 0.36 51.47 36.32 Red Spruce 0.21 0.34 0.33 0.22 22.56 22.69

Models exhibited similarity in regards to variable importance selection. The same five variables were selected for each presence/absence model, though rank varied (Table

2.3). For the abundance models, a total of seven different variables were selected, including the five that were selected for the occurrence models. All selected important variables are interactions. No topographic variables were determined as important in these analyses.

PRMTCM was selected as an important variables for all sixteen models. Histograms of

PRMTCM for each species illustrates the frequency of occurrence relative to the entire study area (Figure 2.4). This indicates that areas where winter precipitation matches or exceeds growing season precipitation and mean temperature in the coldest month is lower than the average of the study area are suitable habitat for the species considered in this analysis.

Balsam fir displays a wider range than the three spruce species. Black spruce’s minimum range approximately matches that of the study area.

The actual plot points and predicted presence/absence objects for each species are presented in Figure 2.5. The absence of a species on a current map does not necessarily ascertain that this species was absent at this location. The mapped prediction objects of the presence/absence models indicate that the models were able to precisely capture species presence, with select instances of overprediction in the Acadian Region. The white spruce model overpredicted presence in interior New Brunswick, but was able to capture populations in northern New England into Canada and along the coast. Specificity for the black spruce model ranked lowest, but model prediction of presence was well maintained for the Acadian Region, capturing distinct populations in northern Maine, along the Acadian coast, and in the northern Adirondacks. The well-defined range of red spruce was captured

Table 2.3. Results of random forest analysis for each species. The prevalence ratio is a ratio of prevalence to an absence sample from within the hypervolume (HV) to an absence sample from outside the HV. OOB = Out of bah; AUC – Area under receiver operator curve; MSE = Mean square error. Prevalen OOB Specifi- Sensiti- Pseudo Species AUC MSE Top 5 Variables ce Ratio Error city vity R2

Presence/Absence

PRDD5, MAPMTCM, Balsam 55-20-25 3.3 94.08 98.84 0.99 - - PRMTCM, MAPDD5, Fire GSPMTCM

PRDD5, PRMTCM, White 50-25-25 4.09 92.40 99.41 0.98 - - MAPMTCM, MAPDD5, Spruce GSPMTCM

MAPDD5, PRMTCM, Black 55-20-25 4.32 91.01 99.49 0.99 - - PRDD5, GSPMTCM, Spruce MAPMTCM

PRDD5, MAPDD5, Red 40-40-20 3.15 95.17 99.37 0.99 - - PRMTCM, MAPMTCM, Spruce GSPMTCM

Presence/Absence with Historical Data

PRDD5, MAPMTCM, Balsam Fir 55-20-25 3.29 94.04 98.89 0.99 - - PRMTCM, GSPMTCM, MAPDD5

PRDD5, MAPMTCM, White 50-25-25 4.05 92.52 99.38 0.98 - - PRMTCM, MAPDD5, Spruce GSPMTCM

PRMTCM, MAPDD5, Black 55-20-25 4.2 91.26 99.52 0.99 - - PRDD5, MAPMTCM, Spruce GSPMTCM

PRDD5, PRMTCM, Red 40-40-20 3.32 94.93 99.31 0.99 - - MAPMTCM, MAPDD5, Spruce GSPMTCM

Relative Basal Area

PRDD5, MAPDD5, Balsam Fir 80-8-12 - - - - 78.35 0.02 PRMTCM, TDMAP, MAPMTCM

PRDD5, MAPDD5, White 80-8-12 - - - - 67.72 0.01 PRMTCM, MAPMTCM, Spruce GSPMTCM

Table 2.3. continued Prevalen OOB Specifi- Sensiti- Pseudo Species AUC MSE Top 5 Variables ce Ratio Error city vity R2

MAPDD5, PRMTCM, Black 80-8-12 - - - - 87.83 0.02 PRDD5, GSPMTCM, Spruce MAPMTCM

GSPMTCM, Red 75-15-10 - - - - 74.67 0.01 MAPMTCM, PRDD5, Spruce PRMTCM, MAPDD5

Relative Stem Density

PRDD5, TDMAP, Balsam Fir 80-8-12 - - - - 71.96 0.03 MAPDD5, TDGSP, PRMTCM

GSPMTCM, White 80-8-12 - - - - 64.96 0.02 MAPMTCM, PRMTCM, Spruce TDMAP, PRDD5

MAPDD5, PRMTCM, Black 80-8-12 - - - - 87.51 0.02 PRDD5, GSPMTCM, Spruce MPMTCM

GSPMTCM, Red 75-15-10 - - - - 73.28 0.03 MAPMTCM, PRMTCM, Spruce TDGSP, MAPDD5

Importance Value

PRDD5, MAPDD5, Balsam Fir 80-8-12 - - - - 76.08 357.40 PRMTCM, TDMAP, MAPMTCM

MAPDD5, PRDD5, White 80-8-12 - - - - 65.32 261.37 PRMTCM, MAPMTCM, Spruce GSPMTCM

MAPDD5, PRDD5, Black 80-8-12 - - - - 88.08 352.60 PRMTCM, GSPMTCM, Spruce MAPMTCM

GSPMTCM, Red 75-15-10 - - - - 73.58 254.55 MAPMTCM, PRDD5, Spruce PRMTCM, MAPDD5

46 a.

Figure 2.3. Density plots for actual versus predicted basal area (a), relative stem density (b), and importance value (c) per species. The density line of the prediction object is overlaid.

Figure 2.3. continued. Density plots for actual versus predicted importance value (c) per species. The density line of the prediction object is overlaid.

by the model including extant populations in southern Appalachia. The balsam fir model was able to capture the wide range of this species.

The mapped prediction objects confirm patterns of underestimation in almost all of the abundance models (Figures 2.6 -2.8). These maps reveal that while exact values were incorrectly estimated, the models were largely able to capture the cline from lesser to greater abundance, particularly for the BA and IV models. Black spruce maps presented the most accurate patterns of abundance, representing populations in Québec and along the coasts of eastern New Brunswick, Nova Scotia, and eastern Newfoundland. Red spruce models portrayed populations in southern Appalachia and concentrations throughout the

Adirondacks into northern New England, New Brunswick, and Nova Scotia. The red spruce

Figure 2.4. Presence versus absence plots' relationship with PRMTCM per species. PRMTCM is the pratio multiplied by the mean temperature in the coldest month (MTCM). Presence plots are represented in white and absence plots in black.

abundance models falsely predicted small populations in coastal Newfoundland and

Labrador. The balsam fir abundance models retained denser populations in

Newfoundland,along the Gulf of St. Lawrence, and along the Appalachian ridge of New

Hampshire into Maine, but missed additional New Hampshire populations. The balsam fir stem density and IV models were able to capture heightened stem count in the

Adirondacks. White spruce abundance was uniformly low for the Acadian Region and the models reflect this pattern while still representing concentrated populations along the Bay of Fundy and the coast of Nova Scotia, as well as pockets in Newfoundland, Anticosti Island, and along the Pennsylvania and New York border.

Results for kappa are displayed in Table 2.4. For the presence/absence models, all values were above the 0.75 threshold, indicating a good to excellent agreement between actual and predicted maps. Kappa was greatest for balsam fir and lowest for black spruce for. Both Khis and Kloc values were high for the models, with all values of Khist above 0.9. This suggests that quantitative categorical similarity, as well as spatial similarity, was highly preserved in model predictions. Kappa values for the abundance metrics confirmed underprediction of actual values. Kappa values were low, falling below the threshold of 0.4, which indicates fair performance. Khist and Kloc were above this threshold signifying a better preservation of categorical and spatial similarity. Balsam fir, overall, exhibited the highest kappa and Khist values for the abundance metrics, while red and white spruce exhibited the highest values of Kloc.

2.4.3. Historical Model Performance

The addition of historical tree data appended 321, 5, 33, and 544 plots, respectively, to the balsam fir, white spruce, black spruce, and red spruce occurrence datasets. The models produced with this additional data were not significantly different than the original presence/absence models in regards to OOB error and AUC measurement (Table 2.3).

Selected important variables were retained between each model, though rank varied.

Current predicted occurrence remains similar, but additional habitats indicated by the

Figure 2.5. Actual and predicted presence for each species. Presence was predicted with and without additional historical data.

51 historical data are represented in southern New Hampshire and western Massachusetts for balsam fir, eastern New York for black spruce, and southeast Massachusetts and

Connecticut for red spruce (Figure 2.5).

2.4.4. Likelihood Model Performance

The likelihood prediction maps reveal a strong correspondence with actual concentrated BA (“actual likelihood”) (Figure 2.8), as well as similarities to the BA predicted output. Black spruce’s likelihood model closely parallels the BA model output, reflecting the goodness of fit for black spruce’s abundance modeling. Red spruce’s likelihood model also exhibits similarity to the abundance model output, capturing pockets of populations in the

Adirondacks, Maine, New Brunswick, and Nova Scotia. Balsam fir and white spruce’s likelihood models are less alike to the predicted BA output than they are to actual likelihoods. Predicted likelihood objects for these two species predict much more suitable habitat than the BA model indicates, though similar hotspots were selected across the landscape. Both the white spruce predicted likelihood and predicted BA output indicate habitat along the Bay of Fundy and Anticosti Island. The white spruce predicted likelihood object additionally specifies northern Maine, northern New Brunswick, western Québec, and Prince Edward Island as additional areas where white spruce was more likely to have suitable habitat. The balsam fir likelihood output indicates a strong possibility for occurrence in western Québec, Maine, and Nova Scotia that were missed by the predicted

BA output. The relationship between the likelihood output and predicted relative BA output are further analyzed in Appendix D.

Spearman’s indicated a strong positive relationship between all likelihood objects and BA abundance models. Average correlation was 0.90, with black spruce BA abundance

52 exhibiting strongest relationship (0.95) with the likelihood object, and red spruce the weakest (0.84). The boxplots in Figure 2.9 also exhibit the relationship between these two variables. A general linear model (GLM) locally weighted scatterplot smoothing (LOWESS) line was added to these graphs to represent a general relationship. The likelihood of presence values range from zero to one for almost all BA output categories greater than zero indicating large deviances from the mean. A correspondence between increasing likelihood and an increase in the mean of each category was exhibited.

2.4.5. Future Predictions of Species’ Distributions

Maps generated for the years 2030, 2060, and 2090 under the ENSEMBLE RCP6 model show shifts north and east in suitable habitat, with the eventual loss of almost all habitat for these species in the U.S. by 2090 (Figure 2.10). In 2030, suitable habitat in the

U.S. was projected to already be limited to northern Maine, New Hampshire, and Vermont, as well as the Adirondacks. White and black spruce habitat was projected to disappear from the U.S. by 2060, though habitat remains in the Acadian Region in northern New Brunswick, the Gaspe Peninsula, and Cape Breton Island, Nova Scotia. Balsam fir and red spruce habitat remains in patches in Maine, New Hampshire, and the Adirondacks. Suitable habitat for balsam fir and red spruce dwindles to only a few high altitude locations along the

Appalachian Mountains in the U. S. by 2090. These include locations in the White Mountains of New Hampshire, and the Longfellow Mountains and Katahdin Mountains of Maine.

Within the Acadian Region, further suitable habitat for balsam fir and red spruce was maintained in the northern and coastal highlands of New Brunswick, as well as Cape Breton

Island. All suitable habitat for white and black spruce was extirpated from the Acadian

Region by 2090. At this time, hotspots for suitable habitat for all four species appear outside

Figure 2.6. Actual and predicted stem density for each species.

Figure 2.7. Actual and predicted importance value (IV) for each species. The same color scale found in Iverson et al. 2008 is used for comparison

Figure 2.8. Actual and predicted basal area (BA) and likelihood model outputs for each species

Figure 2.8. continued. Actual and predicted basal area (BA) and likelihood model outputs for each species

Table 2.4. Kappa values for all models. Kloc = Kappa comparative location measure; Khisto = Kappa comparative histogram measure.

Presence/Absence Basal Area Stem Density Importance Value Balsam Fir Kappa 0.83 0.34 0.29 0.34

Kloc 0.90 0.48 0.44 0.48

Khist 0.93 0.71 0.66 0.70 White Spruce Kappa 0.81 0.30 0.28 0.32

Kloc 0.90 0.50 0.52 0.53

Khist 0.90 0.59 0.55 0.60 Black Spruce Kappa 0.77 0.30 0.30 0.31

Kloc 0.83 0.48 0.47 0.47

Khist 0.94 0.64 0.65 0.65 Red Spruce Kappa 0.80 0.35 0.30 0.32

Kloc 0.87 0.54 0.47 0.50

Khist 0.93 0.65 0.64 0.65

the Acadian Region in Québec along the St. Lawrence River Valley and the Gulf of St.

Lawrence, including the Gaspé Peninsula and Anticosti Island, and in interior and northern

Newfoundland along the northern most reaches of the Appalachian Mountain chain.

While potential habitat was diminished between each period, losses in the U.S. are met with significant gains to the north for balsam fir and white spruce, and to the northeast for red spruce (Table 2.5). Black spruce is likely to occupy regions past the northern extent of the study area used in this analysis. Balsam fir and white spruce will have the greatest area of potential suitable habitat available in 2090. Each species loses area by 2090 when compared to current predicted suitable habitat. Balsam fir (48.0%) and black spruce (72.9%)

58 lose the most, while white and red spruce only experience reductions of 31.2% and 21.1% of suitable habitat.

The inclusion of the historical tree data made significant differences in the predictions of future suitable habitat for all four species considered in this analysis (Figure

2.10, Table 2.5). The predicted habitat for black and red spruce in 2030 revealed additional suitable areas throughout the Adirondacks and northern New York, the Champlain Valley, and western Massachusetts. In 2060, additional habitat in Québec was shown for balsam fir and white spruce. Black and red spruce habitat expanded into the Pennsylvania and New

Figure 2.9. Boxplots exhibiting the relationship between predicted likelihood and predicted relative area abundance for each species. The red line is a locally weighted scatterplot smoothing (LOWESS) line that represents a general relationship between the two objects

York border, the Adirondacks, Vermont, southern New Hampshire and the St. Lawrence

River Valley of Québec. Balsam fir, black spruce, and red spruce all gained additional habitat in 2090 in the U.S., primarily in northern and central Maine, as well as the Adirondacks and the eastern border of Vermont. Predicted suitable habitat for red spruce expanded the most in each time period, followed by black spruce, while white spruce gained the least with the addition of the historical observations.

Future suitable habitats were generated using the likelihood prediction object with the inclusion of the historical data for each species (Figure 2.11). These predictions were in consensus with the presence/absence future maps, but reveal prospective core areas of abundance. No locations with a likelihood greater than 85% were predicted to be within the

U.S. for any of the four species by 2060. Hotspots for future suitable habitat in 2090 with this analysis were similar to those listed above, but were more focused. These hotspots included Island, and the Côte-Nord area along the Gulf of St. Lawrence within Québec.

2.5. Discussion

Modeling alternative dependent variables for different species allowed for the close inspection of the important effects of modeling inputs, as well as inherent species characteristics, on model performance. All presence/absence models yielded excellent statistical results, but these models output generated little information about the prioritization of lands for conservation. The addition of historical data to the overall dataset, indicated the persistence of additional habitat on the southern edge of species’ ranges under different models of climate change, which is promising for the maintenance of current forest composition under different models of climate change. This additional, unique dataset also highlights the drawbacks of modeling species distributions using a single

60 current inventory, and calls into question previous models of this sort that have been used to make management decisions. Statistical results for the abundance models were less accurate, but this is not surprising considering the range of values is infinitely greater for abundance models than the binomial prediction for presence/absence models, and the fine spatial resolution used in this analysis. All abundance models underpredicted actual quantities, but were able to maintain relative patterns of abundance across the landscape, allowing for the qualitative prioritization of land. The likelihood prediction object from the presence/absence models was also able to reflect cores of abundance and displayed similarity to the BA predictions. This is an important interpretation of presence/absence models as they are calculated with more computational ease and from a data type that is typically obtainable at a regional scale

The results of this analysis emphasize the importance of accounting for the role of past land use and other factors on the current realized niche of a species, particularly when developing bioclimatic models. Past work modeling species climatic niches has relied primarily on contemporary inventory data, which does not account for the full climatic classification of a species (Tinner et al, 2013). Known datasets are limited by their inability to capture the fundamental niche of species and species are thought to shift within the range of their fundamental under different climate scenarios and associated changes to the competitive forest environment (Maiorano et al., 2013). While it is difficult to capture these realized range shifts without additional paleoecological research, supplementary historical data typically expands the realized niche, adding to overall knowledge of species specific climatic tolerances. In this study, small extensions of range boundaries that have been obfuscated by settlement and development had large implications on future suitable

Table 2.5. Area (thousands of kilometers) occupied by each species under three different models. The reduction in area from current values is shown below the future values in parenthesis. The four sets of results listed are (clockwise) the presence/absence model, the presence/absence model with historical data, and the likelihood model: area with a greater than 50% likelihood of occurrence and greater than 85% likelihood of occurrence.

Presence/Absence Model Presence/Absence Model (Historical) Species Current 2030 2060 2090 Current 2030 2060 2090

Balsam 1,521 1,302 1,142 791 1,523 1,370 1,220 870 Fir (-14.4) (-24.9) (-48.0) (-10.1) (-19.9) (-42.9)

White 971 941 815 668 950 946 867 713 Spruce (-3.1) (-16.1) (-31.2) (-0.4) (-8.7) (-24.9)

Black 1604 1,005 753 434 1,617 1,033 817 506 Spruce (-37.3) (-53.1) (-72.9) (-36.1) (-49.5) (-68.7)

Red 495 469 401 391 504 525 518 578 Spruce (-5.3) (-19.0) (-21.0) (4.2) (2.8) (14.7)

Likelihood Model: Above 50% Likelihood Model: Above 85% Species Current 2030 2060 2090 Current 2030 2060 2090

Balsam 1,522 1,373 1,222 872 828 392 285 211 Fir (-9.8) (-19.7) (-42.7) (-52.7) (-65.6) (-74.5)

White 973 949 870 715 410 137 140 126 Spruce (-2.5) (-10.6) (-26.5) (-66.6) (-65.9) (-69.3)

Black 1,605 1,035 818 508 1173 492 233 70 Spruce (-35.5) (-49.0) (-68.3) (-58.1) (-80.1) (-94.0)

Red 496 527 521 583 257 107 80 70 Spruce (6.3) (5.0) (17.5) (-58.4) (-68.9) (-72.8)

Figure 2.10. Future predicted presence or absence for each species. Predictions generated in 2030, 2060, 2090 under the ENSEMBLE RCP6 climate scenario with and without historical data.

Figure 2.11. Future predicted likelihood for each species. Predictions generated in 2030, 2060, 2090 under the ENSEMBLE RCP6 climate scenario.

64 climate. For example, historical data indicated populations of black spruce in western

Massachusetts that were otherwise not recorded. This additional habitat extended the maximum range value for PRDD5, resulting in increased predicted suitable habitat in the

U.S. under the historical model. The extension of climatic niches via integration of historical data also suggested a greater level of persistence for each species in the southern portion of the Acadian Region under projected climate change relative to models based solely on contemporary data. This is consistent with work examining Abies alba abundance in central

Europe, which found a lack of future range contraction for this species once historical data had been integrated into climate niche models (Tinner et al., 2013). Such findings underscore the profound implications of relying solely on current species distributions in developing models for informing vulnerability assessments and anticipating climate impacts across the landscape.

The possibility does exists that these supplementary areas and associated habitats have already been affected by climate change. Substantial changes in species composition and spruce habitat are known to have been altered with a 0.55°C change in temperature

(Gajewski, 1988), while temperatures in the Northeast have risen approximately 1°C in the last century, with greater increase along the shoreline from New Jersey to New Hampshire

(Wake et al., 2006). Previous studies suggest that anthropogenic influence has had more effect on species composition shifts in the U.S. than climate change in the Eastern U.S.

(Nowacki and Abrams, 2014). It is likely the truth is an interactive effect between declining climatic suitability and anthropogenic disturbance, particularly on the southernmost edge of indicated ranges. Coastal habitats, including Cape Cod, have long been recognized as refugia locations for spruce species due to their cool climates (Schauffler and Jacobson, 2002), and

65 historical data confirms former occurrence of species in these areas. These absent coastal habitats have possibly been excised by development and disturbance alone. Determining the reason for these habitats’ disappearance though is not as important as recognizing former species-climatic relationships that could bear on future habitat suitability, as even persistence in non-optimal climate ranges is indicative of species’ survival tactics in the face of climate change.

Previous studies have found that variation in model performance is greater among tree species than among techniques (Guisan et al., 2007), and that no technique can rescue species that are difficult to predict. These analyses confirmed this trend, with consistent ranking in model performance amongst the species analyzed here. For presence/absence models, it has been observed that generalist species with a widespread range perform worse than species that occupy specific niches (Guisan et al., 2007; Luoto et al., 2005; Pöyry et al., 2008). All four species’ models performed excellent in regard to AUC, but OOB varied.

Of the four species inspected in this analysis, black spruce had the broadest range of distribution, occupying a widespread variety of environments, while red spruce occupied the most specific niche. These species’ ranges are reflected in their mapped model performance, with the current distribution of red spruce captured very well, while black spruce had the lowest rates of specificity and was likely overpredicted in some areas.

However, it is important to note the differences between AUC scores in this analysis were relatively minimal and all above 0.90.

The results of the abundance models anecdotally appear related to the distribution of relative abundance. For example, a majority of the black spruce data points used in this analysis were concentrated at the northern extent of the study area where relative BA and

66 stem density reach 100%. Abundance modeling performed the best for this species as actual abundance increases along a latitudinal gradient associated with climatic clines. The species with the smallest pseudo R2 values for abundance modeling was white spruce. This species has a high distribution of low abundance across a large range and model performance was erratic. The difficulty in distinguishing patterns of abundance for this species is additionally compounded by uncaptured life history traits. For example, over the last century, white spruce has increased in presence and abundance in coastal area due to its ability to outcompete after farm field abandonment (Mosseler et al., 2003). Similarly, with the advent of the pulpwood industry in the Acadian Region over the last century, balsam fir has increased in abundance. Models are likely capturing a larger portion of this species fundamental niche, which is not realized across the landscape, particularly in undisturbed areas. Finally, red spruce abundance models consistently underpredicted values driving up the difference between actual and predicted means. It is likely that the consistent model performance was due to the small and specific current range of red spruce, which reflects both its narrow ecological niche, as well as anthropogenic activity including selective logging of this species from lower elevation mixedwood stands in the 19th and early 20th centuries

(Kelty and D’Amato, 2006). The restricted distribution of red spruce gives rise to an unusual pattern of relative abundance, where it decreases monotonically, but never achieves great numbers, and likely led to overall underprediction in abundance models.

Hypothesis testing is not relevant to random forest objects (Cutler et al., 2007), but the other metrics of model comparison signaled that the BA abundance models were more harmonious with actual conditions than models with stem density or IV as the dependent variable. BA is often the primary variable in forest modeling as it is reflective of established

67 density and biomass (Li et al., 2011). Meanwhile, a large stem count is not necessarily reflective of suitable habitat, as early stages of stand development often exhibit large numbers that reflect recent recruitment as a product of disturbance and might not reflect long-term patterns of abundance. Extensive works in abundance modeling have made use of IV, which takes into account both BA and stem count, weighting each variable equally

(Iverson et al., 2008). Similar distribution patterns were exhibited between all three variables, current and predicted, across the landscape. In many different spruce-fir forest types, high stem density and small diameters are not uncommon, and it was thought that the IV or stem density models might best capture these environments. Many of these environments though, including recently disturbed sites dominated by balsam fir, poor northern sites occupied by black spruce, or the understory of mature stands with a high red spruce, in the form of advanced regeneration, presence, are simultaneously dominated by the same or other spruce or fir species and are likely captured in the BA model (Seymour,

1992).

The overprediction of low values, and corresponding decrease in mean, exhibited in the abundance outputs does not disregard these models as a useful conservation tool.

Areas with predicted, falsely or otherwise, relative low abundance of an at-risk species are unlikely to be chosen for conservation of critical habitat (Guisan et al., 2013). Of greater concern for conservation decision making is the inability of the abundances model to capture locations with the greatest abundance. Though abundance models did repeatedly underestimate actual BA, stem density and IV, they were able to detect locations of greatest abundance and maintained patterns of density across the landscape. Previous works in abundance modeling have maintained that these models are important as they can reflect

68 the core of distribution (Iverson et al., 2011). The abundance models produced in this analysis achieved this, but the probability object of the presence/absence model also displayed parallel patterns to BA abundance. These likelihood models were not only able to accurately detect areas where species were more abundant, but also simultaneously indicated greater probability of occurrence. This is an important interpretation of these outputs, as species that have low abundance across the landscape (e.g. white spruce) are not likely to perform well with abundance modeling, but the most suitable habitat for that small frequency can be detected. These models are as useful as their abundance counterparts and are superior in regards to the wider availability of presence/absence data, as well as the reduced computational capacity needed to perform this analysis.

Remarkable consistency was exhibited throughout all 20 models in regards to variable selection. Seven total climate variables were selected from the total spread of 43 independent variables. All variables selected were climate interactions, emphasizing the importance of both precipitation and temperature in determining suitable species’ habitats.

PRMTCM was selected in all twenty models, PRDD5, MAPMTCM, and MAPDD5, were selected in nineteen, and GSPMTCM in seventeen. TDGSP and TDMAP were selected one and three times, respectively, in the abundance models. Temperature variables such as

MTCM and DD5 reflect a preference or tolerance for colder climates for all species, while precipitation variables indicate preferences for wet weather concentrated in the winter months. Previous works have emphasized the importance of summer temperature as an indicator of species occurrence and growth (Duveneck et al., 2014; Ribbons, 2014) and have examined the correlation between mean July temperature and the treeline (Cogbill et al.,

1997). On the other hand, recent biogeographical studies suggest that tolerance to climate

69 extremes, particularly freezing temperatures, accounts for 80% of variation in range size

(Mathews and Bonser, 2005; Pither, 2003). MTWM, closely related to July temperature, was not as good of an indicator of species occurrence as cold weather variables, suggesting that tolerance for cold temperatures on a landscape scale was more important than limiting summer maximums for predicting species occurrence.

While habitat for spruce and fir is predicted to vastly shrink in the U.S. and throughout the Acadian Region, results suggest that extensive areas of suitable habitat will persist in Canada. Hotspots include the Gaspé Peninsula and other high elevation areas along the Gulf of St. Lawrence, Anticosti Island, and interior and northern regions in

Newfoundland. Small populations along the Appalachian Mountains in Maine and New

Hampshire will be important locations for refugia in the United States. These predicted locations of absence and refugia are in agreement with similar analyses for the “boreal conifer forest” under future climate scenarios with the added beneficial effects of increased carbon dioxide (CO2) (Tang and Beckage, 2010), though other studies have stated that increased CO2 would have little to no effect on growth of spruce as the optimum temperature for photosynthesis is exceeded (Ollinger et al., 2007). Coastal habitats were not predicted as important locations in future predictions, with the exception of red spruce in Nova Scotia. Pollen records shows that white and black spruce were able to persist in coastal New England during a period of warming between 6000 and 5000 years due to cool and foggy refugia habitat generated by tidal mixing in the Bay of Fundy (Schauffler and

Jacobson, 2002). Future climate projections though predict that the Bay of Fundy and the

Gulf of St. Lawrence will warm (Bush et al., 2014) and downscaled sampling of global models predict increased temperatures along the coast at rates at least equal to nearby inland

70 habitats and neglect to mention large scale changes in regional ocean circulation (Hayhoe et al., 2008).

Red and white spruce saw smaller reductions in area of suitable habitat than balsam fir and black spruce, and habitat for red spruce and balsam fir was predicted to persist in the

U.S. until 2090. The smaller percent reduction of red and white spruce habitat is primarily due to their current restricted range sizes, and the persistence of red spruce and balsam fir habitat is because these species are more tolerant of warmer temperatures than white or black spruce. Red spruce is projected as being restricted to high elevations areas within the

United States, as well as in Nova Scotia, New Brunswick, and Québec. This is harmonious with general beliefs of the effects of climate change on plant species (Parmesan and Yohe,

2003), but also in-line with red spruce life history traits. This includes exemplified patterns of adaptability, currently maintaining habitat at suitable elevations in Appalachian

Mountains of West Virginia, as well as rapidly establishing itself in the Acadian Region only between 1000 and 500 years ago when temperatures cooled by 1°C (Lindbladh et al., 2003).

Current high elevation habitat that is predicted to persist under different models of climate change, should be prioritized for conservation. Warmer temperatures will likely increase growth in red spruce habitat that is currently surviving at the edge of its cold tolerance, such as krummholz and other diffuse form Acadian high altitude environments (Gamache and

Payette, 2004; Harsch et al., 2009). While balsam fir is predicted to lose a substantial amount of area, the total predicted suitable habitat for this species is greater than any other species considered in this analysis, as the models were able to detect a larger share of this species’ fundamental niche. Due to this species’ comparative tolerance for warmer temperatures, large range, high abundance, high fecundity, and competitive superiority in

71 disturbed environments, the future outlook for suitable realized habitat under climate change is positive (Bakuzis and Hansen, 1965). The reduction in black spruce habitat was relatively exaggerated as the study area was unable to show the complete northern range of this species. As a plastic species, tolerant of a wide range of climate and soil conditions, controls on black spruce habitat is influenced less by abiotic features and more by biotic competition (Murphy et al., 2006). Persistence in current habitat in the Acadian Region, such as the vast complexes of peatlands in the lowland of Maine, will rely on the maintenance of current hydrology and the delay in arrival of swamp competitors, such as black tupelo (Nyssa sylvatica), whose northern edge of ranges’ are currently restricted by cold temperatures. While the maintenance of black spruce ecosystems in the Acadian

Region might be difficult, the persistence of this species is likely in the core of its range in

Canada. Lastly, white spruce, is extremely restricted in its current range in the U.S., and is easily outcompeted in poor light and soil environments. While white spruce has the phenotypic plasticity to be able to survive in a variety of climate conditions (Gordon, 1996), it suffers from the ability to adequately migrate due to its restricted current range. This species is a good candidate for current habitat protection and facilitated migration through the establishment of populations in proper suitable habitat in northeastern Canada.

Basic tenets of forest management are built around historical forest conditions and requirements of individual species and ecological sustainability; the assumption being that if we maintain these conditions, the forest will continue to provide goods and services. With impending shifts of species distributions due to climate change these assumptions can no longer be held true. Creating models that can accurately link species distribution with climate is an essential first step towards visualizing future landscapes as climate warms.

These models are static predictors of species’ dynamic with climate and with one another, and cannot take into account species specific phenotypic plasticity, longevity, fecundity, or dispersal that will determine migratory success in a competitive environment as climate changes. Rather by exploring the impacts of different data and dependent variables inputs, these models assist in delineating and ranking future suitable habitat for conservation, and elucidating species specific patterns across the landscape.

2.6. Conclusion

Presence/absence models generated with the random forest algorithm were able to precisely predict species occurrence on the landscape. Both abundance models and the likelihood prediction object from the presence/absence models represented the core distribution of the important species considered in this analysis. The likelihood object was easy to generate and its interpretation allows land managers to determine the most likely habitat for species of concern, particularly important for species of inherently low abundance. The inclusion of historical data in these analyses was important in elucidating habitats that have been excised by anthropogenic disturbance and species’ habitats were predicted to persist further south in their range under climate change. Future habitats for spruce and fir species will be sparse in the U.S., limited to sections of the Appalachian

Mountain chain in Maine, New Hampshire, New York, and Vermont. Suitable habitat was projected to be present to the north and east in Canada, located on interior and peninsular

Newfoundland and along the Gulf of St. Lawrence in Québec, including the northeastern tip of the Gaspé Peninsula, the Côte-Nord region, and Anticosti Island. The models created in this analysis were reliable and can be used to inform current and future management decisions.

CHAPTER 3

MODELING AND FORECASTING THE INFLUENCE OF CURRENT AND FUTURE CLIMATE ON

MAXIMUM STAND DENSITY FOR EASTERN NORTH AMERICAN SPRUCE-FIR FORESTS

3.1. Abstract

The spruce-fir (Picea-Abies) forest type of the Acadian Region is at risk of disappearing from the United States and parts of Canada due to climate change and associated impacts. This valuable ecosystem provides habitat to wildlife of both local and national conservation concern, and sustains regional economies. Managing for the many ecosystem services provided by this forest type requires accurate forecasting of forest metrics across this broad international region in the face of the expected redistribution of tree species. The maximum stand density index (SDImax) has long been used by foresters to determine stocking potential and phases of stand development based upon a stand’s species composition and location. Previous predictions of the SDImax for spruce-fir forest types were limited by specific-species composition mixtures that could not be applied outside of the study area and use of traditional statistical methods. Using linear quantile mixed models (LQMM), predictions of SDImax were readily estimated for spruce or fir- dominated plots across the Acadian Region. Model performance was strong and estimates of SDImax from these models were similar to previous regional studies. Estimated slope coefficients for the relationship between quadratic mean diameter and stand density ranged from -1.46 ± 0.21 (± SE) for white spruce (Picea glauca (Moench) Voss) to -2.20 ±

0.11 for balsam fir (Abies balsamea L.), varying from Reineke’s universal slope of -1.605 .The establishment of an individual constant slope of self-thinning for plots dominated by each spruce or fir species reinforces previous research that Reineke’s slope is not universal for all

74 species, and that the differences in slope are telling of different species’ life history patterns. Providing regional estimates based on the estimation of plot species dominance, obviates the necessity for the construction of specific species ratios, and simplifies the estimation of potential SDImax. Individual plot estimates of SDImax, achieved through a varying intercept, allowed for assessment of each stand’s potential and limitations, and for a wide range of inferences about the impact that climate, nutrient availability, site quality, and other factors might have on a stand’s SDI.

Estimates of SDImax for each species were linked with climate and topographic variables using the nonparametric random forest algorithm to generate models that accurately predicted changes in species’ SDImax under different models of climate change. This represents the first known formal analyses of region wide differences in species specific

SDImax due to climate, though previous research suggests that climate does have an influence on stand stockability. Model performance was consistently high (average pseudo

R2 of 84.3), and the random forest models were able to approximate the observed regional patterns of SDImax, though very low values were consistently overpredicted. Varied climate variables were selected for each species’ model, consistent with known specific species climatic requirements. The spatial distribution of spruce-fir forest types’ SDImax under the

ENSEMBLE RCP6 climate show a general pattern of shifts in SDImax values to the north and east over the next century with the almost complete extirpation of these species, and their associated SDImax, in the U.S. by 2090. While the mean SDImax is expected to decrease on average for all species, this reduction remains steady over the next century, and similar maximum SDImax values are achieved elsewhere on the landscape as species’ distributions shift.

3.2. Introduction

In the last three decades, global temperatures have increased more so than they have in any other 30-year period over the last 1400 years. Additionally, global surface temperatures are expected to rise by another 0.3 - 4.8°C by the end of the 21st century

(Stocker et al., 2013). It is already widely recognized that as the climate warms, many species migrate poleward and upward, with one recent analysis of mobile organisms finding a median latitudinal migration of 16.9 km per decade (n=764) having already occurred due to climate change (Chen et al., 2011). For sessile organisms, such as forest trees, rapid migration potential is limited, and shifts in the suitability of habitat conditions (Iverson et al.

2008), or the reconfiguration of forest structure (Dolanc et al., 2013), is a more common outcome of climate warming. Of particular concern in the United States (U.S) is the spruce- fir (Picea-Abies) forest type in the Acadian Region. It already realized that this ecosystem is at risk from disappearing from the U.S. due to climate change, with previous climate analyses predicting range contraction of up to 400 km north (Iverson et al., 2008) and a possible reduction of 97-100% of suitable habitat in the U.S. in the next 100 years (Hansen et al., 2001).This valuable ecosystem provides habitat to wildlife of both local (e.g.,

Canadian Lynx (Lynx canadensis), spruce grouse (Dendragapus canadensis canace)) and national conservation concern (e.g., Bicknell’s Thrush (Catharus bicknelli)), and sustains regional economies (McWilliams et al., 2005). Managing for the multiple resources provided by this ecosystem requires accurate forecasting of the relationship between climate and important forest metrics through easily applied, flexible modeling techniques in the face of the redistribution of species’ habitats. Previous species-climate models have focused on presence/absence, which while useful for the spatial distribution of future habitat, it does

76 not yield the information necessary to make informed forest management decisions in regards to forest production and carbon storage.

Density management diagrams (DMDs), which graphically represent the relationship between average tree size and stand density in forests, have long served as an important tool in making predictions about future stand development based on forest management decisions (Jack and Long, 1996). Imperative to designing DMDs are the concepts of stand- density index (SDI; Reineke, 1933) and relative density (RD; Drew and Flewelling, 1977).

Both of these indices are comparative measurements that provide the degree to which a stand is achieving full site occupancy based upon the maximum size-density relationship

(SDImax) (Zeide, 2005). SDI is defined as “the number of trees per hectare as if quadratic mean diameter of the stand is 25 cm” (Long, 1985) and is calculated using the slope of the

SDImax line, while RD is ratio of observed SDI to a species- and region-specific SDImax. The

SDImax is part of a linear continuum demonstrated for a diversity of plant life forms, the self- thinning line, where a stand with a few large trees or one with many small trees, fall on either end, and along this continuum a stand will self-thin due to competition at a constant rate (Yoda et al., 1963).

Traditionally, the SDImax has been determined through visual observations of the most fully stocked stands, and all other stands of similar species composition are compared to this SDImax to obtain their SDI and RD. The RD of a stand generally corresponds to important phases in stand development (Drew and Flewelling, 1977). For example, at a RD of ~.15 (phase I), crown closure is obtained, between a RD of .35 and .55 the growth of individual trees slows (phase II), and above .55 a stand enters the “zone of eminent

77 mortality” and asymptotically approaches the SDImax self-thinning slope (phase III) (Hann,

2014). Applying these principles allows for a compromise between the maximization of production at the stand level and the maximization of individual tree growth (Long, 1985).

Additionally, other important forestry concepts, such as stockability, or the tolerance of a forest stand to the presence or competition of an increasing number of trees, are generally inferred from the SDImax (DeBell et al., 1989). In the early life stages of stands, increased productivity would be associated with faster growth rates, but as a stand enters phase III, an increase in a stand’s stockability accounts for greater productivity. Thus, establishing a species’ SDImax and obtaining relative measures such as RD or SDI are imperative to applying appropriate management techniques based on the stand life history stage and managing for desirable future forest conditions. In order for this tool to be functional for managing future forests, the relationship between size-density patterns and climate needs to be fully realized.

Numerous equations have been proposed to define the SDImax across stands of different species compositions (Drew and Flewelling, 1977; Reineke, 1933; Yoda et al.,

1963). While the size variables differ (i.e. volume, quadratic mean diameter (QMD), height, crown size) between these formulations, these equations have in common the intent to define the slope that describes the self-thinning line. The completeness of these equations has been called into question over the last thirty years (Volvfovicz Leon, 2011). In theory, the intercept for each of these equations vary for stands of different species composition, the slope is constant, and the SDImax achieved is independent of site index, age, and management as it is commonly assumed that these additional factors only influence the time it takes for a stand to reach the maximum (Jack and Long, 1996). Previous studies have

78 argued that not only does the species-specific intercept vary based on factors such as site nutrient quality (Morris and Myerscough, 1991), soil fertilization (Bi, 2004), site index

(Weiskittel et al., 2009), and stand age (Zeide, 2005), but also that the slope is not universal across all species types, and that each stand or population has its own dynamic thinning line

(VanderSchaaf and Burkhart, 2007; Weiskittel et al., 2009; Zeide, 1987). Given the potential sensitivity of these relationships to site conditions, the effects of current and future climate conditions on size-density relationships needs to be explored.

Regardless of the ongoing debate about the commonality of species-specific intercepts and the slope of the self-thinning line, a major hindrance of the maximum size- density line is its inability to account for structurally diverse and mixed-species stands, where a near infinite number of species combinations, and corresponding varying intercepts, occur. Early solutions for structurally diverse stands included calculating the SDI through a summation method, where the SDI is calculated for each tree or diameter class individually, and summarized to yield stand SDI (Shaw, 2000; Stage, 1968). While this is able to account for diverse stand structures, it does not account for varying species compositions. Various techniques have been explored to estimate SDImax of mixed species stands and typically fall into one of three categories: (1) stands of a mixed-species forest type are selected and a static SDImax boundary line is developed through different statistical techniques (i.e. modified linear regression (Solomon and Zhang, 2002; Sturtevant et al.,

1998; Williams, 2003), reduced major axis regression (Wilson et al., 1999), fixed effects nonlinear regression (Yang and Titus, 2002)); (2) stands of a mixed-species forest type are selected and a dynamic SDImax surface that varies with species composition is developed through different statistical techniques (i.e. fixed effects nonlinear regression (Puettmann et

79 al., 1992), modified linear and non-linear regression (Stout and Nyland, 1986; Swift et al.,

2007), optimization functions (Rivoire and Le Moguedec, 2012)); or (3) SDImax is calculated and statistically related to a proxy for varying species distributions such as specific gravity

(Ducey and Knapp, 2010; Woodall et al., 2005)) or top height (Sterba and Monserud, 1970).

While many of these techniques have been successful in approximating SDImax for their datasets, they suffer from the inability to easily be extrapolated to other regions, as well as complicated model forms.

Quantile regression is a statistical method that has recently been introduced to ecological studies and has proven itself as an effective tool for modeling the SDImax (Cade and Guo, 2000; Zhang et al., 2013, 2005). Quantile regression involves inspecting the relationship between two variables at quantiles of the distribution other than the mean, the standard in linear regression. This is particularly useful in evaluating heterodastic datasets with unequal variances, where multiple rates of changes are distributed through different quantiles (Cade and Noon, 2003; Koenker and Bassett Jr., 1978). Additionally, quantile regression eliminates the need to subjectively select plots that have already achieved their

SDImax (Zhang et al., 2005). Quantile regression has been used in a variety of ecological studies to study the upper boundary of a relationship between two variables, establishing the effects of a constraint on a response (Lima-Ribeiro et al., 2014; Niinemets and

Valladares, 2006; Stahl et al., 2014). The constraining relationship between plant density and plant size has been examined in multiple studies (Cade and Guo, 2000; Sea and Hanan,

2012) including, specifically, SDImax in forestry applications (Cao and Dean, 2015; Zhang et al., 2013, 2005).

Quantile regression though has been criticized for the difficulty to make statistical inference from the results (Zhang et al., 2005), as there is no defined distribution on the error portion of the model (Cade and Noon, 2003). The deterministic component of the model is parametric and is defined as the inverse of the cumulative distribution function of the response variable (Koenker and Bassett Jr., 1978). Estimates of prediction have been achieved without any parameterization on the error (Cade et al., 1999), as well as goodness of fit measures (Koenker and Machado, 1999), but have come under scrutiny by the authors themselves later in time. Additionally quantile regression in its simple linear form is unable to account for variation both within and amongst plots whether it be due to established modifiers (i.e. species composition) or those that are more widely debated (i.e. site quality and stand origin). Previous applications of quantile regression in SDI studies, have either used SDI as the dependent variable and predicted variations due to mean stand variables

(Ducey and Knapp, 2010; Woodall et al., 2005) or have incorporated these variables directly into the equation (Zhang et al., 2013).

Linear quantile mixed models (LQMM) apply quantile regression methods to mixed models, allowing for varying intercepts as well as varying slopes. Developing models for clustered data (e.g. measurements taken at the same plot at different times) at points other than the mean has been discussed over for the last two decades (Jung, 1996; Koenker,

2004), but only recently has a model been developed that can estimate fixed and random effects at any quantile (Geraci and Bottai, 2007). The LQMM developed by Geraci and Bottai

(2007) does not follow the inverse of cumulative distribution function of simple quantile regression, but rather the model is based on the asymmetric Laplace distribution (ALD). ALD has been suggested for use in quantile regression before, as likelihood ratios tests (Koenker

81 and Machado, 1999) and variance of the distribution (Yu and Zhang, 2005) are more easily defined. Additionally, as with other mixed models, clustered data is accounted for, and variation at the individual and group level is calculated (Jones, 2007). The data is partially pooled and groups with even only one observation can be predicted and provide information to the overall estimation of coefficients and variance (Gelman and Hill, 2006).

The popularity of SDImax can be attributed to the fact that it is grounded in plant population biology theory (Reineke, 1933), and its wide applicability to a diversity of ecosystems, including different forest types. This near universal concept is an important tool for managing current, and potentially, future forests when properly applied. All forests are likely to experience changes in their distribution and productivity due to climate change

(Dolanc et al., 2013; Iverson and Prasad, 1998; Medlyn et al., 2011; Mohan et al., 2009)).

Many studies have analyzed the relationship between site differences and SDImax, usually finding that site index, which in turn is influenced by climate, is linked to stand variability

(Weiskittel et al., 2009, 2011; Zhang et al., 2013). Debell et al. (1989) and Harms et al. (1994) observed that stands of loblolly pine (Pinus taeda) in Hawaii and South Carolina, experienced large differences in stockability without increased mortality, and assigned the differences to the more favorable climate and nutrient availability in Hawaii, as well as differences in sunlight angle. Not one study though has explicitly studied the regional relationship between SDImax and climate, and how this relationship might change under different models of climate change. Inspecting the relationship between climate and species at a landscape level seeks to model the species-boundary line, or the maximum size-density of a given species across all environments and populations (Weller, 1990). Recent implementation of classification and regression trees (CARTs) such as the random forest

82 algorithm (Liaw and Wiener, 2002), make it possible to study the effects of a mass array of climatic variables on a dependent variable. Random forest has proven excellent at selecting the most important climatic variables from a multi-variable space where the relationships between the response and predictors is not always linear, and has the ability to predict this relationship onto future climate landscapes (Joyce and Rehfeldt, 2013).

The overall intent of this study is to develop an easily applied technique that can model SDImax for specific species in mixed species forests, in order to study the relationship between SDImax and climate at a landscape level. The specific objectives of this analysis are to: (1) use linear quantile mixed effects modeling to estimate the maximum size density for spruce/fir (Picea-Abies) dominant plots of the northeast; (2) determine the importance of climate and other factors on estimating the SDImax using random forest; and (3) predict the future distribution SDImax under different models of climate change.

3.3. Methods

3.3.1. Data

A total of 10,493,619 tree observations on 248,821 plots were considered to determine the SDImax of spruce-fir forest types across the study area. The data collection period spanned from 1955 to 2012 and the majority of the data was collected after 1980

(85%). The New Brunswick Department of Natural Resources, the Newfoundland Forest

Service, the Nova Scotia Department of Natural Resource, the Québec Ministry of Natural

Resource, and the Prince Edward Island Department of Agriculture and Forestry provided coverage of Canada. In the U.S., the Cooperative Forestry Research Unit, the Massachusetts

Department of Conservation and Recreation, the National Park Service, the New Hampshire

Division of Parks and Lands, the U.S. Forest Service (USFS), the University of Maine, the

University of Massachusetts, the Vermont Center for Ecostudies, and the Vermont

Monitoring Cooperative provided data.

Plots with a dominant spruce or fir species component were determined from the data. The four species selected for were balsam fir (BF; Abies balsamea (L.) Mill.), white spruce (WS; Picea glauca (Moench) Voss), black spruce (BS; Picea mariana (Mill.) B.S.P.), and red spruce (RS; Picea rubens Sarg.). The data was cleaned prior to analysis. For plots with two or more measurements, observations were removed that had not yet reached the phase of competition induced mortality (i.e., RD < 0.55). Additionally, data in the 99th percentile, that were thought to also be in phase 1 or phase II, were removed prior to

LQMM analysis, to properly estimate the intercept and slope of the SDImax self-thinning line.

3.3.1.1. Climate Data

Climate data was collected from Moscow Forest Science Laboratory climate database available online at http://forest.moscowfsl.wsu.edu/climate/ (download date 05

January 2014). Climate data was derived by applying thin-plate smoothing spline procedures that extrapolate data from discrete weather stations to specific plot points with corresponding elevation (Rehfeldt, 2006). Current climate data was normalized for a thirty year period (1960-1990) and was based on weather station data for about 15,000 locations for precipitation and 12,000 for temperature (Joyce and Rehfeldt, 2013). 33 climatic variables were used in analysis, sixteen of which are direct measurement of climate, while the remaining seventeen are interactions (Table 3.1).

3.3.1.2. Topographic Data

Topographic variables were used to model species occurrence and abundance in order to capture discrete landscape features that influence a species’ dynamics and life history outcomes, and also to capture effects that terrain features might have on microclimate. Elevation, slope, and aspect data were collected, if available, from the original data source. If not available, elevation data was extracted from the 30 m resolution national elevation dataset (NED) generated by the United States Geological Service (USGS) available at http://viewer.nationalmap.gov/viewer/ (download date 12 February 2013) and from the

30 m resolution digital elevation dataset made available through the Canadian Council on

Geomatics (CCOG) available at http://www.geobase.ca/geobase/en/find.do?produit=cded

(download date 3 March 2014). Slope and aspect were derived from the NED using the raster package (Hijmans, 2014) available through R statistical software (R Core Team, 2013).

A measure of northness and eastness were calculated from aspect data based on Beers et al. (1966). Five additional topographic indices were derived from the NED using the System for Automated Geoscientific Analyses (SAGA) (Brenning, 2008), including a topographic wetness index, a convergence index, a terrain index, a topographic openness index, and site curvature. These variables were assumed to capture effects not reflected in the climate variables such as soil drainage, exposure, and solar radiation profiles.

3.3.2. LQMM Analysis

The purpose of the first phase of this analysis was to estimate the maximum size- density index (SDImax) for plots with a dominant spruce or fir component in the Acadian

Region. Reineke’s (1933) SDI equation was selected due to the widespread availability of the

Table 3.1. Description of climate variables used in this analysis. Mean, standard deviation (S.D.), minimum (Min), and maximum (Max) values are listed for both the plots used in this analysis and the entire study area. Climate variables in bold represent those which were used to construct the absence sampling hypervolume. Plots Study Area Acronym Definition Mean S.D. Min Max Mean S.D. Min Max Julian date of when the number D100 106.0 40.1 15.0 188.0 114.8 40.6 17.0 197.0 of days above 5°C reaches 100 Annual number of days below 0°C DD0 based on mean 735.5 673.1 0.0 3233.0 975.8 898.7 0.0 3480.0 monthly temperature Annual days above 5°C based 1222. DD5 2491.7 1234.3 503.0 5431.0 2267.2 356.0 5372.0 on mean monthly 4 temperature Julian date of first freezing FDAY 278.4 21.7 238.0 347.0 274.1 22.2 237.0 349.0 temperature in autumn Frost free period FFP 150.7 48.8 59.0 297.0 141.9 49.0 58.0 298.0 length Mean number of days above 5°C 1084. GSDD5 2067.0 1096.7 311.0 4994.0 1883.5 240.0 4858.0 between SDAY 3 and FDAY Growing season (April - GSP 611.0 71.5 396.0 1109.0 588.5 70.0 323.0 1128.0 September) precipitation Mean annual MAP 1185.5 177.1 656.0 2217.0 1110.9 188.9 654.0 2374.0 precipitation Mean annual MAT 8.5 5.8 -5.2 19.9 7.0 6.4 -6.4 19.7 temperature Mean maximum temperature in MMAX 26.8 4.2 14.1 33.9 25.9 4.8 12.4 33.9 the warmest month Mean minimum MMIN temperature in -10.8 7.8 -32.2 4.9 -13.0 9.0 -32.2 4.7 the coldest month Annual number of days below 0°C based on mean 1226. MINDD0 1367.2 964.7 13.0 4690.0 1685.2 19.0 4943.0 minimum 1 monthly temperature Mean MTCM temperature in -5.1 7.9 -24.6 11.4 -7.4 8.9 -25.5 10.8 the coldest month

Table 3.1. continued Plots Study Area Acronym Definition Mean S.D. Min Max Mean S.D. Min Max Mean temperature in MTWM 20.8 4.0 10.4 27.8 20.0 4.5 8.9 27.6 the warmest month Julian date of last freezing SDAY 127.0 26.8 49.0 184.0 131.5 26.9 52.0 187.0 temperature in spring TDIFF MTWM-MTCM 25.9 4.5 15.8 37.0 27.3 5.0 16.6 37.2 Interactions Annual dryess ADI index: 0.0 0.0 0.0 0.1 0.04 0.01 0.01 0.06 (DD5)0.5/MAP Annual dryness & ADIMINDD cold index: ADI * 50.0 32.3 0.6 176.4 61.9 40.7 1.1 182.4 0 MINDD0 DD5MTC (DD5 * -3.5 18.7 -26.0 61.9 -6.5 15.5 -26.5 58.0 M MTCM)/1000 GSPDD5 (GSP * DD5)/1000 1560.7 888.3 240.4 4777.3 1383.1 853.5 198.8 4234.4 GSPMTC (GSP * -2.9 4.8 -15.9 9.5 -4.0 5.0 -17.1 8.5 M MTCM)/1000 GSPTD (GSP * TDIFF)/100 156.8 24.7 96.0 282.7 158.8 24.5 80.8 295.7 1731. MAPDD5 (MAP * DD5)/1000 3023.5 1716.8 514.8 8787.9 2649.3 427.7 8289.3 4 MAPMTC (MAP * -5.4 9.0 -28.9 16.4 -7.0 9.0 -36.1 14.4 M MTCM)/1000 (MAP * MAPTD 302.3 42.8 191.0 508.8 296.4 36.2 179.5 593.5 TDIFF)/100 MTCMGS MTCM/GSP 0.0 0.0 0.0 0.0 0.0 0.0 -0.1 0.0 P MTCMMA MTCM/MAP 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 P PRATIO GSP/MAP 0.5 0.0 0.4 0.7 0.5 0.0 0.4 0.7 PRDD5 PRATIO * DD5 1295.5 657.7 227.7 3392.3 1196.0 618.1 166.0 3267.0 PRMTCM PRATIO * MTCM -2.7 4.3 -15.8 7.1 -4.2 5.1 -16.1 6.6 Summer dryness SDI index: 0.1 0.0 0.0 0.1 0.1 0.0 0.0 0.1 (GSDD5)0.5/GSP TDGSP TDIFF/GSP 0.0 0.0 0.0 0.1 0.0 0.0 0.0 0.1 TDMAP TDIFF/MAP 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1

87 size variable (mean tree size) over a large study area, the ability of this variable to be adjusted to reflect a wide variety of stand structures, and the flexibility to objectively select stands that are at or near the SDImax. Reineke (1933) defined the relationship between size and density as:

푙푛 (푇푃퐻 ) = −1.605 푙푛 (퐷 ) + 푧 where D is average stand diameter, 푧 is a constant varying with species, and -1.605 is the slope Reineke (1933) estimated that explains the relationship of self-thinning between density and size. From Reineke’s (1933) formula, the SDI for any given stand can be calculated as:

푆퐷퐼 = 푇푃퐻( 퐷 ) 1.6 25.4

TPH and mean tree size were determined for each plot in this study. Reineke’s

(1933) diameter (DR) was used as an alternative to the traditional QMD, in order to account for the varied and non-normally distributed tree sizes often found in the spruce-fir forest types of the U.S. DR is calculated as:

1 1 퐷 = ( ∗ ∑ 퐷푖1.6)1.6 푅 푇푃퐻

where 퐷푖 is an individual tree diameter, that is summed per stand. Using DR in the SDI formula yields the same result per stand as the summation method often used for calculating SDI in unevenly distributed stands (Shaw, 2000; Zeide, 1983).

The ln-ln relationship between TPH and DR for each species was modeled using the

LQMM package (Geraci, 2014) available in R (R Core Team, 2013). Individual plots were

88 accounted for through a random effect on the intercept, while the slope coefficient was fixed. From the results of the LQMM model, individual plots’ ln TPH was estimated. SDImax was calculated using the fixed slope coefficient from the output of the model in place of

Reineke’s (1933) slope.

LQMM is similar to linear quantile regression, in that the quantile function of the response variable is the inverse of the cumulative distribution function. Given a sample of observations, (훸푖, 푦푖), where 훸푖 is a matrix of predictors for each value of 푦푖, the quantile

−1 function in linear quantile regression is 푄푦푖훸푖 = 퐹푦푖훸푖. The goal is to estimate the quantile, τ, by:

휏 푄푦푖|훸푖(휏) = 훸푖훽 , 푖 = 1, … , 푁,

휏 휏 where τ is 0 < 휏 < 1 and 훽 is a matrix of coefficients for 훸푖 at 휏. 훽 is estimated by minimizing a loss function of absolute values of residuals, defined as 𝜌휏(푣) = 푣(휏 − 퐼 (푣 ≤

0) ), where 퐼 is an indicator function. This loss function assigns weights of τ or 1 - τ to observations based on whether they are greater or lesser than the mean, meaning that

푃푟 (푦 ≤ 휇) = 휏 (Geraci and Bottai, 2007). Errors change as a function of 훸푖, but there is no assumed distribution, and the variance of the regression is volatile, changing with small differences in τ (Zhang et al., 2005).

For LQMM, where a sample of observations is (훸푖푗, 푦푖푗) and each 푖 is nested within a

푗 푄 = 퐹−1 group, , the quantile is estimated as 푦푖푗|푘푖 푦푖푗|푘푖 where the response is conditional on a location-shift random effect, 푘푖, that is independently distributed according to the ALD

(푦푖푗 ~ 퐴퐿퐷 (휇, 𝜎, 휏)). Thus LQMM the linear mixed quantile model of the response is written as:

푄푦푖푗|푘푖(휏|훸푖푗, 푘푖) = 훸푖푗훽 + 푘푖, 푗 = 1, … , 푛푖, 푖 = 1, … , 푁,

and 푦푖푗 conditional on 푘푖 is distributed as:

휏(1−휏) 1 푓(휇, 𝜎, 휏) = 푒푥푝 {− 𝜌 (푦 − 휇 )}, 휎 휎 휏 푖푗 푖푗

where 휇푖푗 = 푋푖푗훽 + 푘푖,−∞ < 휇 < ∞ acts as the location parameter, 𝜎 > 0 acts as the scale parameter, and 0 < 휏 < 1 acts as the skewness parameter (Geraci and Bottai, 2007; Geraci,

2014; Yu and Zhang, 2005) 𝜌휏 is the loss function defined above. The mean and variance of

1−2휏 휎2(1−2휏+2휏2) the ALD (푦 ~ 퐴퐿퐷 (휇, 𝜎, 휏)) are 훦(푦 ) = 휇 + 𝜎 and 푉퐴푅(푦 ) = , 푖푗 푖푗 휏(1−휏) 푖푗 (1−휏)2휏2 respectively, and are proofed in Yu and Zhang (2005).

3.3.3. Random Forest Analysis

Once SDImax for each plot was estimated using LQMM, the relationship between

SDImax and climate was inspected using the randomForest (Liaw and Wiener, 2002) package in R. Random forest is a type of CART that is able to predict the response variable by creating multiple trees that select predictors that minimize error, and then aggregating the results of these trees to determine output. Each tree is generated by sub-sampling two- thirds of the complete data set and then recursively partitioning the data by choosing the optimal predictor variable for splitting the data at each node. Random forest is unique in that at each node a subset of the independent variables are selected. This added layer of randomness reduces correlation between trees and thus decreases total forest error rate

(Breiman, 2001). Additionally, selecting from a subset of independent variables increases computational efficiency making this algorithm ideal for large datasets with a multiple dimension independent variable space. Partitioning is complete once the error can no

90 longer be reduced and multiple terminal nodes are reached. The result is a tree that predicts for the dependent variable at each terminal node, by means of deriving the average response value in from the observations within this node using a piecewise constant prediction function (Strobl et al., 2009).

The random forest algorithm reserves one-third of the model dataset, referred to as the out of bag (OOB) sample, for each tree that is constructed. This sample is used to internally estimate the precision of the tree constructed by running the sample down the tree and recording the accuracy of each data point’s value (Breiman, 2001). For regression trees, the mean square error (MSE) as well as a “pseudo R2” is calculated and reported to determine accuracy. Random forest’s R2 differs from the traditional R2 in that the variance is calculated by dividing by n, as opposed to n-1.

The number of spruce or fir SDImax points were duplicated prior to random forest analysis. Increasing prevalence within a dataset relative to the actual incidence across the landscape decreases erroneous errors of absence without violating any basic statistical assumptions (Pearson and Dawson, 2003). Additionally points where a dominant spruce or fir species was known to be absent, were appended to the input dataset, in order to train the model to distinguish between areas with similar abiotic features but with dissimilar species composition and SDImax. Preliminary analyses showed a ratio of 90 to 10, occurrence to absence, provided the most accurate results. Half of the absence data were sampled from areas determined to be climatically similar to areas where spruce-fir forest types were present. To establish climatic similarity, an eighteen variable hypervolume (Table 3.1, in bold) was defined and expanded by 0.01 standard deviation in all directions (Joyce and

Rehfeldt, 2013). Absence data from the hypervolume were stratified by ecoregion and randomly sampled. To complete the dataset, additional “outside” absence data were randomly sampled from beyond this established hypervolume.

A random forest consisting of 500 trees was run five times. Using the VarImp function option in the random forest package, the most important variables were determined using the unscaled permutation accuracy importance measure for each forest.

This measure is a calculation of the mean decrease in node impurity, when a variable in the tree is randomly permutated to another variable. Permuted variables which result in a higher decrease in purity are considered more important. The unscaled computation of this measure was used as the scaled measure and has shown preference of correlated predicted variables (Strobl et al., 2007) with these results provided greater predictive accuracy.

Preliminary analyses showed that iteratively reducing the complete array of 43 variables to the 5 most important variables resulted in a model that retained model accuracy while parsimonious balancing computation efficiency and an accurate description of SDImax. Final models were generated using the most important independent variables in a random forest with 500 trees.

3.3.4. Current and Future Predictions

Mapped predictions of future distribution of the SDImax for spruce and fir forest types of the Acadian Region were generated using the output of the random forest predicted over different climate landscapes in the years 2030, 2060, and 2090. All mapping was based on 0.00833° (~1 km2) grid and generated with the raster package (Hijmans, 2014) in R. Future landscapes were acquired for each important variable through the Moscow

Forest Science Laboratory’s climate database. The ENSEMBLE representative concentration

92 pathways 6 (RCP6) scenario, generated in affiliation with the International Panel on Climate

Change (IPCC), was used to forecast future suitable habitat. These RPCs were created by analyzing varying predicted rates of radiative forcing, as well as greenhouse gases emission rates and concentrations by the year 2100 (Stocker et al., 2013). RCP6 is a moderate scenario, the 6 referring to the radiative forcing in 2100 measured in watts per square meter.

3.4. Results

Of the 248,821 total plots considered in this study, 80,133 were found suitable for analysis. A total of 15,143 or 18.9% of these plots were classified as predominantly one of the four spruce or fir species. The majority (79.9%) of the spruce or fir plots had only one measurement, 7.1% had two measurements, and the remaining 13.0% had three or more.

The majority of the spruce or fir plots were classified as predominantly BF (49.5%; Figure

3.1, Table 3.2). The observed TPH for balsam fir and black spruce was considerably larger than white or red spruce, and DR was lowest for these two species. In general, plot climate data reached the minimum and maximum values of the study area and the means of the plots were within one standard deviation of the means of study area (Table 3.1).

The relationships between the observed ln(TPH) and ln(DR) suggested that individual plots viewed collectively across the landscape would yield a self-thinning trend line, and that these lines would vary by the dominant species selected (Figure 3.2). Predicted slopes

(±S.E.) from the LQMM models ranged from -2.20 (±0.11) for balsam fir to -1.46 (±0.21) for white spruce (Table 3.3). All predicted intercepts and slopes were significant and plot level random effects for the intercept had a large range. The relationships between the predicted

Figure 3.1. Map of different spruce-fir forest types distributed across the study area. Plots are colored according to their forest-type. BF = balsam fir; BS = black spruce; RS= red spruce; and WS = white spruce.

Table 3.2. Summary of stand variables for each species. TPH = trees per hectare, DR = Reineke’s diameter, Comp % = the percent of composition that the species occupies on the plot, and S.D. = standard deviation. Dominant Species No. of Variable Mean S.D. Minimum Maximum Plots TPH 2003.0 2629.0 59.5 24980.0

Balsam Fir 7288 DR 14.3 4.6 4.5 33.5 Comp % 64.4 20.6 19.7 100.0 TPH 1076.0 989.5 74.3 12430.0 White Spruce 843 DR 16.9 5.3 4.5 38.3 Comp % 62.0 22.3 20.9 100.0 TPH 1749.0 2137.2 75.0 24480.0 Black Spruce 5626 DR 12.4 3.5 4.5 30.2 Comp % 82.1 18.6 21.1 100.0 TPH 1304.0 1180.7 59.5 14640.0

Red Spruce 1668 DR 17.7 4.6 4.5 38.1 Comp % 61.2 19.5 16.8 100.0

94 ln(TPH)99 and observed ln(DR) are consistent with the expectation that for the every value of

DR the TPH prediction would increase to represent the 99th quantile and the maximum ln(TPH)99 and observed ln(DR) are consistent with the expectation that for the every value of

DR the TPH prediction would increase to represent the 99th quantile and the maximum relationship between the two variables along a consistent slope (Figure 3.3). The mean values of predicted TPH were largest for balsam fir and black spruce and much lower for the

99th percentile group for all species and for two species, white and black spruce, white, and red spruce (Table 3.4), consistent with the observed TPH pattern (Table 3.2).

Correspondingly, the mean predicted values of SDImax were lowest for balsam fir and black spruce, and their predicted slope lines the steepest.

Figure 3.2. Observed ln (TPH) vs observed ln (DR) for all plots used in this analysis. The average lqmm trend line for each species is overlaid. TPH = trees per hectare; DR = Reineke’s diameter.

Summary statistics were calculated for those plots considered to be in the 99th percentile and are presented in Table 3.5 in order to compare the difference between the

th 99 percentile plots and the total plots, per species. The mean DR between the two subsets are very similar. The percent of total composition for each species was on average higher for the 99th percentile subset, but safely within one standard deviation of the mean of total plots and not considered significantly different. The mean TPH was consistently higher in exceeded one standard deviation in difference from the mean of the total plots.

The average pseudo R2 (± MSE) of the random forest object developed using the complete array of 43 climate and topographic variables was 84.9 (± 9341.7). Reducing this display of variables to the 8 most important, decreased the R2 to 84.0 (± 10830.1), and to the 5 most important, decreased the R2 to 83.24 (± 11435.4). The random forest model for white spruce was the least accurate and displayed the most variability (79.2 (± 18453.2)), while the red spruce model achieved the highest R2 (91.8 (± 9212.6)) and the black spruce model achieved the lowest error (80.1 (± 7771.9) (Table 3.6). The random forest models were able to closely match the spatial distribution of the species across the landscape, as well as gradients in SDImax quantities (Figure 3.4). White spruce’s presence, as well as low

Table 3.3. Statistics of predicted intercept and slope for each species’ linear quantile mixed model. Dominant Standard Variable Value Lower Bound Upper Bound Species Error Intercept 13.59 0.29 13.00 14.17 Balsam Fir Slope -2.20 0.11 -2.41 -1.98 Intercept 11.49 0.54 10.41 12.57 White Spruce Slope -1.46 0.21 -1.88 -1.04 Intercept 13.06 .47 12.12 14.01 Black Spruce Slope -2.07 .17 -2.42 -1.73 Intercept 12.38 0.21 11.96 12.81 Red Spruce Slope -1.76 0.08 -1.92 -1.60

96 values of red spruce, were overestimated, including the inaccurate presence of low quantities of red spruce in Newfoundland. Overestimation of low values are responsible for drawing down the predicted means of SDImax, particularly red spruce, though overall the models were able to match the distribution of mid-range and higher values (Table 3.7). The average ratio between the predicted and actual means was 0.32, but if only predicted values over their observed minimum are considered, this ratio rises to 0.80.

The top 5 most influential variables for all models were all climatic and primarily interactions. Three temperature variables (DD0, DD5, and D100) appeared throughout the spruce models, signaling the importance of temperature in influencing the limits of these species’ SDImax. For example, as DD0, or the number of days where the temperature is

Figure 3.3. Predicted ln (TPH) vs. observed ln (DR) for all plots in this analysis. The average lqmm trend line for each species is overlaid. TPH = trees per hectare; DR = Reineke’s diameter.

97 below 0°C, increases, black spruce’s SDImax increases, and no values are found below 500

(Figure 3.5). Black spruce SDImax values also decreased as PRATIO increased, indicating preference for an environment where precipitation is concentrated in the winter months.

As D100, or the Julian date of when the days above 5°C reaches 100, increases, the SDImax of white spruce peaks and then decreases, signifying less tolerance for colder weather. White spruce also exhibited a preference for moister summer weather. PRDD5 was a top two important predictor in two models (balsam fir and white spruce) and both species’ SDImax show a significant drop as values increase, reflecting an intolerance for hotter temperatures, particularly balsam fir. Some important climatic variables, such as MTCMGSP and TDMAP in the red spruce model, exhibited a normal distribution for SDImax values, reflecting the core likely distribution of red spruce’s values, and the fact that its entire range was captured in this model.

The spatial distribution of spruce-fir forest types’ SDImax under the ENSEMBLE RCP6 climate show a general pattern of shifts in SDImax values to the north and east over the next century with the almost complete extirpation of these species, and their associated SDImax, in the U.S. by 2090 (Figure 3.6). While the mean SDImax is expected to decrease on average

10.4% for all species, this reduction remains steady over the next century, and similar maximum SDImax values are achieved elsewhere on the landscape as species’ distributions shift. The reduction of balsam fir and red spruce SDImax values are more gradual than white or black spruce, with these species persisting in select portions of their range in the U.S., though diminished, until 2090. SDImax values of balsam fir are predicted to increase in

Newfoundland and will expand into new territory in interior western Québec, though SDImax might be limited in this region. The SDImax of red spruce will gradually decrease in most of

Table 3.4. Summary of predicted variables for each species using linear quantile mixed models (LQMM). TPH = predicted trees per hectare, SDImax = calculated stand density index maximum, and S.D. = standard deviation. Dominant Variable Mean S.D. Minimum Maximum Species SDImax 683.4 124.0 175.4 1258.0 Balsam Fir TPH 3888.6 4470.1 299.8 34347.4 SDImax 911.9 123.9 572.9 1238.1 White Spruce TPH 2038.8 1319.3 451.8 10426.2 SDImax 601.3 85.2 322.5 894.1 Black Spruce TPH 3614.2 3250.8 425.7 29024.7 SDImax 852.8 226.7 176.7 2075.0 Red Spruce TPH 1992.0 1596.8 190.8 18111.9

Table 3.5. Summary of stand variables for each species in the 99th percentile. TPH = trees per hectare, DR = Reineke’s diameter, Comp % = the percent of species composition that the dominant species occupies on the plot, and S.D. = standard deviation. No. of Dominant Species Variable Mean S.D. Minimum Maximum Plots TPH 4553.0 4774.5 825.0 22980.0 Balsam Fir 291 DR 14.4 4.5 4.6 26.1 Comp % 77.4 18.8 25.8 100.0 TPH 2444.0 2092.3 850.0 11290.0 White Spruce 40 DR 16.8 5.3 5.3 26.9 Comp % 68.1 23.9 28.8 100.0 TPH 4564.0 4338.9 1175.0 24230.0

Black Spruce 191 DR 12.5 3.4 4.5 21.2 Comp % 90.4 13.3 28.2 100.0 TPH 2325.0 2005.8 625.0 14600.0 Red Spruce 73 DR 17.9 4.8 4.6 31.8 Comp % 70.1 18.7 30.0 98.9

Table 3.6. Results of the SDImax random forest models for each species. Important variables are listed in order of their importance. SDImax = maximum stand density index ; M.S.E = mean square error. Dominant Species Psuedo R2 M.S.E. Important Variables

Balsam Fir 83.5 9232.6 PRDD5, DD5MTCM, PRMTCM, GSPDD5, ADIMINDD0

White Spruce 79.2 18453.2 PRDD5, D100, GSPDD5, SDI, DD0

Black Spruce 80.1 7771.9 DD0, PRMTCM, GSPMTCM, PRATIO, MAPDD5

Red Spruce 91.8 9212.6 MTCMGSP, TDMAP, DD5MTCM, PRDD5, DD5

Figure 3.4. Maps of current actual versus predicted SDImax. Actual SDImax generated with the linear quantile mixed model (LQMM) versus the random forest prediction of SDImax based on climate.

100

Nova Scotia over the next century, but will increase in northern New Brunswick, the Gaspé

Penninsula in Québec, and on Cape Breton Island in Nova Scotia. Red spruce habitat and larger SDImax values are predicted to expand further into Québec, along the northern coast, the St. Lawrence River Valley and Anticosti Island, as well as Newfoundland. Major reductions in habitat are predicted for white and black spruce as early as 2030 in the southern portion of their range, and little to no habitat is represented in the U.S., except in northern Maine, at this time. Reductions of white spruce’s SDImax values are not as severe in

Table 3.7. Summary of SDImax predicted using the random forest models for each species. The ENSMEBLE RCP 6 climate scenario was used for predictions in 2030, 2060, and 2090. Values in parenthesis in the current column represent the ratio between the listed value and the SDImax generated with the linear quantile mixed models (Table 3.4).Values in parenthesis in the future columns represent the ratio between the listed value and the current prediction in the first columns. Since the random forest models suffered from overprediction of low values, all minimums were set to the minimums in Table 3.4. SDImax = maximum stand density index. S.D. = standard deviation Future Dominant Current Species 2030 Mean S.D. Max Mean S.D. Max Balsam 558.8 133.4 1029.0 480.4 147.5 954.9 Fir (0.82) (1.08) (0.82) (0.86) (1.11) (0.93) 740.8 101.4 1066.1 725.3 89.4 1025.1 White Spruce (0.81) (0.82) (0.86) (0.98) (0.88) (0.96) 529.7 71.3 835.8 489.8 77.2 749.5 Black Spruce (0.88) (0.84) (0.93) (0.92) (1.08) (0.90) 581.9 224.0 2070.6 523.5 194.0 1131.2 Red Spruce (0.68) (0.99) (1.00) (0.90) (0.87) (0.55) Future Future Dominant 2060 2090 Species Mean S.D. Max Mean S.D. Max Balsam 458.4 150.1 962.9 471.8 151.7 870.5 Fir (0.82) (1.13) (0.94) (0.84) (1.14) (0.85) 718.0 89.3 1039.8 711.6 100.4 1007.1 White Spruce (0.97) (0.88) (0.98) (0.96) (0.99) (0.94) 471.9 75.9 745.5 476.0 80.6 731.7 Black Spruce (0.89) (1.06) (0.89) (0.90) (1.13) (0.88) 492.0 190.4 1111.5 506.1 208.1 1345.7 Red Spruce (0.85) (0.85) (0.54) (0.87) (0.93) (0.65)

101 the remaining habitat, and new suitable habitat with high predicted values are expected in interior central Québec and eastern and peninsular Newfoundland. Black spruce’s SDImax values will continue to diminish over the next century in most of the study area, though new growth is predicted for Cape Breton Island until 2060, and peninsular Newfoundland. Black spruce will likely extend into territories north of the study area considered in this study.

Overall, the models indicate spruce-fir populations in the U.S. will be severely restricted to

Figure 3.5. Partial dependency plots for each species’ random forest model and the two most important variables from those models. The most important variable is listed on the x- axis, the second most important variable on the y-axis, and the predicted SDImax value is list on the z-axis.

102 high elevation areas, particularly in Maine, though populations will persist throughout the

Acadian Region in Nova Scotia, and in coastal and northern New Brunswick.

3.5. Discussion

Modeling for stands dominated by specific species in a mixed species landscape using LQMM successfully provided species boundary lines, as well as individual plot estimates of SDImax. The error component of the LQMM algorithm (Geraci, 2014; Yu and

Zhang, 2005) afforded effective gauging of error and the range of uncertainty for predictions, as opposed to fixed effects quantile regression (Zhang et al., 2005). The establishment of an individual constant slope of self-thinning for plots dominated by each spruce or fir species reinforces previous research that Reineke’s slope is not universal for all species (Pretzch and Biber, 2005; Weiskittel et al., 2009), and that the differences in slope are telling of different species’ life history patterns. The emphasis of this study was not on how species composition might change the SDImax, as the effects of species composition on

SDImax is known to be highly species-composition specific (Woodall et al., 2005), but rather to find if a specific species dominance was an important enough factor to establish a distinct and constant slope of self-thinning. Previous studies of mixed-species stands were only able to account for populations of a limited geographic scope and specific species-composition ratios, which has relatively little context outside of their intended study area (Solomon and

Zhang, 2002; Stout and Nyland, 1986; Sturtevant et al., 1998). Providing regional estimates based on the estimation of plot species dominance, obviates the necessity for the construction of specific species ratios, and simplifies the estimation of potential SDImax.

Individual plot estimates of SDImax, achieved through a varying intercept, allowed for

103

Figure 3.6. Maps of future predictions of SDImax depicted as a ratio between the future predicted value and the current predicted value. Deep red indicates a sharp decrease, while deep green represents expansion of species; SDImax into new territory. Models were generated using random forest models for each species under the ENSEMBLE RCP6 scenario of climate change in years 2030, 2060, and 2090.

104 assessment of each stand’s potential and limitations, and for a wide range of inferences about the impact that climate, nutrient availability, site quality, and other factors might have on a plot’s SDI. Due to the breadth of the input data, representing diverse stand conditions across a large study area, analysis of the effects of regional drivers, including climate, was straightforward. Climate was found to be an important determinant in

2 establishing patterns of SDImax, for each species across the landscape, with psuedo R values ranging from 79.2 for white spruce to 91.8 for red spruce. Information provided from this study can be used to plan for future conditions that will arise as the growth and distribution of species migrate due to climate change.

Species’ individual inherent life history traits, as well as anthropogenic activities on the landscape, influence self-thinning trends, which in turn affect the predicted SDImax, and random forest model behavior. Anecdotally, it appears that species with higher density values in this study (i.e. balsam fir, black spruce) and lower average size values, mathematically resulting in lower values of SDImax, and a steeper slope line. In the case of balsam fir, this species has increased in dominance in the Acadian Region as an early successional competitor due to aggressive harvesting, including salvage logging as a reaction to the late 1970s spruce-budworm outbreaks in the region. The steep slope exhibited in this study might be a by-product of a forest in transformation, as the stages of succession are passed through and balsam fir’s presence on the landscape is reduced (McWilliams et al.,

2005). Additionally, both balsam fir and black spruce dominated stands tend to form unabated by competition, often in full light, and young stands consist of a high abundance of small trees. These species are also able to grow on poor sites, such as thin-soiled montane locations for balsam fir (Sprugel and Bormann, 1981) and the Canadian shield and

105 lowlands for black spruce (Subedi and Sharma, 2013), resulting in mature stands with relatively high TPH and small BA. Unsurprisingly, stands of this nature might thin faster and exhibit qualities similar to that of shade intolerant species. Meanwhile, red spruce often grows in the understory as advanced regeneration before being released and coming to dominate a stand. By the time a red spruce stand is calculated as dominant based upon basal area, the stand has already likely already self-thinned to some degree, resulting in larger average size values and a shallower slope. White spruce, which is naturally sparse and inconsistent across the landscape, exhibited the most variation in both the lqmm and random forest model and is difficult to account for. Naturally this species grows in the moist cool fog belt of the Acadian coast, and often co-exists in mixed stands. White spruce has increased in dominance across the landscape due to its ability to thrive in farm fields abandoned over the last century, though it is outcompeted over time (Mosseler et al.,

2003). It is likely that the LQMM model was shaped around these older semi-natural stands, resulting in a shallower slope.

Validation of results through direct comparisons between previous predictions of

SDImax for the species in this study are difficult, as earlier studies inspected specific mixed species composition ratios. For example, Sturtevant et al. (1998) studied mixed balsam fir- black spruce-miscellaneous stands in Newfoundland, with an average ratio composition of

74-17-9, and found a SDImax of 1050. Subsetting the data for similar location and species composition conditions, the LQMM predicted SDImax ranging from 379 - 950, with a mean of

760. Similarly, Swift et al. (2007) estimated the SDImax for 50-50 spruce-fir mixtures in New

Brunswick as 900. The LQMM predictions for similar conditions range from 435-1190 with a mean of 844. Although the ratios used in Solomon and Zhang (2002) are unclear, SDImax

106 predictions ranged from 992 for spruce-fir as well as hemlock (Tsuga canadensis)-red spruce mixtures to 1310 for cedar (Thuja occidentalis)-black spruce mixtures, which are comparable to SDImax estimates for the species in this study.

Comparing the slope of the self-thinning lines is also difficult, as earlier studies of similar species used different formulas or independent and dependent variables. The 95% confidence interval for the slope coefficients predicted using LQMM encompasses Reineke’s

(1933) slope of -1.68 for both red and white spruce. Meanwhile the predicted slope coefficients for both black spruce (-2.07±.08) and balsam fir (-2.20±0.11) are steeper than

Reineke’s value and the upper and lower bounds do not encompass -1.68. Both the

Solomon and Zhang (2002) and Wilson (1998) formulations for spruce-fir plots in Maine, as well as the Swift et al. (2007) study of the Acadian Region, found slopes shallower than the -

3/2 power law when developing a self-thinning line based on ln(Max Volume)-ln(TPH) relationship. However, steeper slope values were reported for the cedar-black spruce forest types in Solomon and Zhang (2002), as well as the values reported in Newton and Smith

(1990) and Newton (2006) based upon the size-density relationship. Slopes similar to the values found in this study for black spruce and balsam fir have been reported, but have traditionally been associated with shade intolerant species, particularly pines (i.e. Pinus contorta, P. echinata, P. elliottii, P. taeda) (Reineke, 1933; Woodall et al., 2005). However, more recent studies have found a wide range in slope predictions for a variety of species, varying far from Reineke’s (1933) established slope. For example, Pretzch and Biber (2005) calculated slopes ranging from -1.204 to -2.027 in mixed stand of beech (Fagus sylvatica L),

Norway spruce (Picea abies [L.] Karst.), Scots pine (Pinus sylvestris L.) and common oak

(Quercus patraea [Mattuschka] Liebl.). In addition, values of -0.593 to -1.687 were reported

107 for mixed stands including white spruce, lodgepole pine, and trembling aspen (Populus tremuloides Michx) (Yang and Titus, 2002).

Steps were taken to alleviate common pitfalls of SDImax modeling in the first phase of analysis. One recent review of multiple SDImax analyses, concluded that modeling the boundary lines of individual stands using datasets with multiple measurements consistently taken over time, and using a size measure (i.e. DR, volume) as the dependent variable, yields the most reliable results (Hann, 2014). While this review was generated from the perspective of stand level growth and trajectory modeling, and this paper is more related to high resolution landscape level patterns due to climate, the Hann (2014) finding does bring up important inconsistencies common in SDImax modeling. For example, concerns about the effect of “meaningless observations” or areas where slope is either infinite or 0, in phase I, before competition induced mortality, can be primarily be dismissed, as inspecting only the

99th quantile automatically excludes the majority of these observations. The data in this analysis was minimally screened to eliminate observations that appeared to be both in the

99th quantile and in phase I of stand development, as not to dampen the coefficient predictions. Additionally, while time-series data is valuable when seeking to model the individual stands trajectories of populations, it is not as valuable when seeking to model a species boundary line, particularly when using a mixed-model approach, where individual and group effects both exert influence on the final predictions. With the exception of this minimal cleaning, the selection of plots used in this analysis was an objective process including a massive dataset encompassing much of the Acadian Region. It is believed that this data does represent landscape wide specific species patterns of self-thinning in a variety of conditions.

108

Another concern is that the 99th quantile of data used in these analyses is inherently different than the remaining data, and predictions on this quantile should not be applied outside of this group. To inspect if this was the case, metrics of TPH, DR, and species composition ratios were compared between the two groups. The mean percent of the species composition ratio of the dominant species between the two groups was not significantly different, eliminating the possibility that differences between the two groups was due to species composition. DR was also similar between the two groups, as expected.

TPH differed between two groups, significantly for white and black spruce. This variation of

TPH between the two groups was expected though, as the ability to have a higher density

th while maintaining the average DR, is what differentiates these observations into the 99 percentile, and establishes the SDImax for a particular species.

The random forest models largely overpredicted the presence of low SDImax values, which are unrealistic in the context of total plot SDImax values. For example, the red spruce random forest model exhibited the greatest problem with the overprediction of low values and underprediction of very high values. Since the entirety of red spruce’s range was captured in the study area, and red spruce exhibits an “abundant-core distribution”, with core habitat surrounding the center and extending 60-70% of the way towards the edge of its range (Murphy et al., 2006), it is believed that the random forest model tried to exact a normal distribution across the landscape. By virtue, the calculation of SDImax for a total plot does not vary normally from zero to a maximum value, as opposed to calculating TPH or BA for only a specific species. This normal absence of low SDImax values, could explain the difference between observed and predicted distributions of SDImax values in all the random

109 forest models. This focus on higher values and species dominance, could lead to more accurate estimation and selection of important climate variables.

When compared to species-specific abundance variables, SDImax consistently showed the highest association with climate variables and resulted in better model accuracy (Thesis,

Chapter 2), with the exception of black spruce models, where the R2 for abundance variables was slightly higher. This is likely due to the large consistent expanses of consistent

TPH, BA, and IV, associated with black spruce, which are easier to detect, as opposed to the relatively more difficult SDImax gradient. Consistent with Chapter 2, the majority of these climate variables were interactions, emphasizing the importance of both precipitation and temperature in determining suitable species’ habitats, though there was a slight preference in the selection of temperature variables. In contradiction to Chapter 2, these variables varied greatly between species’ models, suggesting that using SDImax as the dependent variable is better at capturing important species specific climate signals. By and large the selection of variables in this study match with known criteria of species’ specific tolerances and preferences. The selection of variables for the black spruce model indicated a steep tolerance for cold weather and for a climate where the majority of precipitation occurs in winter (Vincent, 1965). While black spruce can survive in warmer environs, it tolerates and thrives in locations of extreme cold, unabated by competition (Pither, 2003). White spruce, also considered a cold tolerant species, was clearly less tolerant of extreme cold than black spruce, and was limited by warming temperatures in the southern portion of its range.

Values of SDImax for white spruce increased with summer moisture, and white spruce is known for its reliance on coastal moisture in this region, particularly fog. While fog or fog drip is not directly captured in this model, it is possible the climate variable SDI, or summer

110 dryness index, captured these moist coastal zones where fog occurrence is greatest in summer months (Klemm et al., 1994). Red spruce has a small specific range, and is known to be reliant on cool moist environments (Dumais and Prévost, 2007). The three climate- precipitation interaction variables selected for the red spruce models indicate such, with a normal distribution of SDImax values in a small constrained range, with steep drop offs in

SDImax outside of this distribution, indicating higher temperatures or limiting moisture. The patterns of SDImax for the two temperature variables selected (i.e. DD5, DD5MTCM) indicate a cold preference. Lastly, though not as cold tolerant as black and white spruce, balsam fir is a generalist with the ability to survive in a wide array of climate conditions (Bakuzis and

Hansen, 1965), and the patterns of the selected variables indicate as much. It is clear though that this species has strict limits in terms of SDImax values, limited by the lack of adequate moisture as well as hot and cold temperatures.

Estimations of SDImax in 2030, 2060, and 2090 represent the potential achievable SDImax for stands that are predominantly composed of the species in this study. While differences in SDImax due to climate are accounted for by the climate models, other factors not captured by this model are certain to play a role in the actual stocking of species in the future.

Previous studies have found that nutrient availability, soils, and angle of the sun all play an important role in stocking, and that greater precipitation appears to affect stockability differences versus temperature, and climate more so than soils (DeBell et al., 1989; Perala et al., 1999). Species specific features need to be taken into account in regards to their influence on realized SDI and stockability. For example, the ability to reallocate foliage along the bole is an individual species trait that affects stockability (Dean and Long, 1992), and clumping, which might be expected in lower light conditions, could lead to lower stocking

111 levels (Puettmann et al., 1993). This study is an important step in estimating potential SDImax for specific species in a mixed species landscape, and estimating shifts in species’ SDImax due to climate.

Managing for future forest stands will not only require knowledge of specific species’ life history requirements in a mixed species landscape, but also an understanding of ecophysiological responses to climate change in different life history stages. Increased heat stress and evapotranspiration, and decreased snowpack and soil moisture, will certainly affect vulnerable seedling recruitment and survival (Nitschke and Innes, 2008), particularly in the case of shallow rooted and moisture dependent white and red spruce seedlings

(Davis, 1966). Drought, changes in nutrient availability, and increased vulnerability to disturbances such as fire and disease as a result of climate change, are thought to be a major driver of mortality in mature stands (Allen et al., 2010). Studies have shown that despite the mortality of mature individuals and seedling stress, forests are not shifting as fast as anticipated, as younger, smaller individuals readily replace the overstory (Dolanc et al. 2013). This suggests that while forest structure is certainly shifting, composition changes are not as certain.

Dynamic ecosystems require dynamic management plans, and frequently adjusted models of species’ abundance across the landscape are essential to informing these plans.

As climate change takes effect, realized niches will shift (Maiorano et al., 2013), phenotypic plasticity will be expressed (Joyce and Rehfeldt, 2013), new species interactions will emerge

(Williams et al. 2013), and species will adapt through migration and by changing structure.

Forest managers should focus now on cornerstones of adaptive forest management by

112 increasing resistance and resilience of forest stands (Noss, 2001). This requires not only the passive protection of forests, including primary forests and areas known to be climatic refugia for species from previous climatic events (Keppel et al., 2012), but also active management. Low intensity forestry, including partial cuts, are more likely to increase resistance and resilience than aggressive forestry practices, as species diversity is maintained and oft increased, soil structure is preserved, and the ecosystem is not left vulnerable to invasives (Noss, 2001). Due to the mixed species nature of the Acadian Forest, as well as the passage of the Maine Forest Practices Act (MFPA) and similar legislation which heavily regulated clearcuts, partial harvests including selection and shelterwood cutting already compose a significant portion of harvesting activity in the Northeast U.S.

(McWilliams et al. 2003), while clearcuts are more prominent in Canada. Aggressive partial harvesting (i.e. Under MFPA only a BA of needs to be maintained) though can also open up forests to risks associated with low species and structural diversity, as well as soil disturbance (Sader et al., 2003). Multi-aged management systems, which mimic natural disturbance, and increase resilience through greater structural and functional diversity, are already being researched in the Northeast (Nunery and Keeton, 2010; Saunders et al., 2008) and are seen as the best management practice in the face of an uncertain future (O’Hara and Ramage, 2013). Furthermore, density management has been suggested as the most effective approach to managing forests for both resistance and resilience to climate change

(Chmura et al., 2011). Using DMDs, constructed from the SDImax values and the self-thinning lines presented in this study, forest density can be reduced to delay the onset of mortality due to drought (Elkin et al., 2015) as well as nutrient stress caused by competition, fire risk, and the predisposition to disease outbreak, all while maintaining economic profitability. For

113 example, by reducing aggressive harvests, while managing for stand density and composition, balsam fir on the landscape would be reduced, along with the increased risks and impacts of a spruce budworm outbreak (Westveld, 1946). Additionally, shifts in forest structure due to climate change can be accounted for with landscape level species maximum boundary lines that inherently account for a diversity of stand structures which aren’t as predominant today, but might be in the future.

3.6. Conclusion

Predictions of SDImax for plots dominated by a spruce or fir species in the Acadian

Region were successfully modeled using LQMM. The establishment of an individual constant slope of self-thinning for plots dominated by each spruce or fir species reinforces previous research that Reineke’s slope is not universal for all species, and that the differences in slope are telling of different species’ life history patterns. Providing regional estimates based on the estimation of plot species dominance, obviates the necessity for the construction of specific species ratios, and simplifies the estimation of potential SDImax.

Individual plot estimates of SDImax, achieved through a varying intercept, allowed for assessment of each stand’s potential and limitations, and for a wide range of inferences about the impact that climate, nutrient availability, site quality, and other factors might have on a plot’s SDI. Climate was found to be an important determinant in establishing patterns of SDImax, for each species across the landscape. Overprediction of low SDImax values were present, thought to be an artefact of the modeling technique. This overprediction of low values is not seen as a concern, as land managers should focus on conserving areas with high potential SDImax. Information provided from this study can be

114 used to plan for future conditions that will arise as the growth and distribution of species migrate due to climate change.

115

CHAPTER 4

CONCLUSION AND SUMMARY OF MAJOR FINDINGS

Climate Change (IPCC), mean global temperatures are predicted to rise between 1.5°C and

4.5°C by 2090 (Stocker et al., 2013). Already, records indicate that temperatures have risen by 0.89°C since 1880. As temperatures rise, cascading changes to the global climate system are taking effect, including transformations to precipitation, humidity, and cloud cover.

These changes to the global climate system are reflected in species’ geographic distributions and ecosystems’ configurations, as each species has a specific set of climatic requirements and limitations that determine their fundamental niche on the landscape. The fundamental niche is bounded by additional abiotic controls, as well as biotic competition, where species compete for requirements including light, nutrients, and water, resulting in the realized niche, often represented as current species distribution. While species of the spruce-fir forest type exist in distinct associations with one another today, paleoecology studies indicate that past compositions have no bearing on current, and likely future, forest assemblages (Davis, 1976).

This thesis was an investigation into the effects of the use of different dependent variables in species distribution models (SDMs) on not only characterizing the relationship between species and climate, but also investigating the difference in model outcomes in regards to conservation management utility. The focus of this climate study was not the process by which we arrive at future landscapes, but rather to envision what future tree

116 distributions might look like under different climate scenarios. The decision of which dependent variable to use in species distribution modeling is based upon the desired management product, where passive management, through the conservation of suitable lands, or the active management of forests, is pursued. Both presence/absence and abundance variables seek to help land managers select the best land for conservation in the face of shifting species distributions due to climate change. Presence/absence models are easier to generate and to interpret, while abundance variables help to pinpoint locations of greater habitat suitability. Alternatively, predicting maximum stand density index (SDImax), allows for the construction of density management diagrams (DMDs), which have long served as an important tool in making predictions about future stand development based on size-density relationships. These different dependent variables were analyzed for spruce

(Picea spp.) and fir (Abies spp.) across the entire Acadian landscape and compared, while also exploring innovative modeling techniques.

4.1. Summary of Findings by Objective

4.1.1. To Explore New Data and Modeling Techniques for SDMs

Previous studies that have predicted range contraction of the spruce-fir forest type have been limited by the absence of data that fully characterizes the species’ relationships with the environment in the northern portion of their range, as it reaches across international boundaries, preventing range wide modeling and monitoring. These previous analyses only made use of a single national inventory, widely thought to underrepresent the

Northeastern spruce-fir resource. Additionally, a course resolution of 20 km2 was used in previous regional climate-envelope analyses of spruce and fir species (Iverson et al., 2008), not allowing for the evaluation of specific lands for future conservation. Lastly, while

117 abundance variables have been considered for this region, alternatives such as SDImax have not. Innovate modeling techniques included the prediction of SDImax on plots dominated by a spruce or fir species using linear quantile mixed models (LQMM).

10,493,619 observations on 248,821 plots from twenty-two different agencies were collected to provide details about the contemporary distribution of spruce and fir species.

Additionally, 1,342 historical tree observations on 778 plots were obtained from a database maintained by Charles Cogbill (Cogbill, 2000). Forest Inventory and Analysis (FIA) data contributed less than 3% to the total observations of spruce or fir, marking a departure from previous analyses which solely relied on FIA data. This data was able to fully characterize the range of the important spruce-fir species within the Acadian Forest, including important locations in Canada, where suitable habitat will exist in the future. The extension of climatic niches via integration of historical data also suggested a greater level of persistence for each species in the southern portion of the Acadian Region under projected climate change relative to models based solely on contemporary data. The high spatial resolution (1 km²) used in this analysis allow for specification of future habitat to the stand level. Models developed using this data resulted in high accuracy and performance, particularly the presence/absence and SDImax models.

The use of species-specific SDImax represents the first known formal analyses of region wide differences due to climate. Modeling for stands dominated by specific species in a mixed species landscape using LQMM successfully provided species boundary lines, as well as individual plot estimates of SDImax. The establishment of an individual constant slope of self-thinning for plots dominated by each spruce or fir species reinforced previous research that Reineke’s slope is not universal for all species (Pretzch and Biber, 2005;

118

Reineke, 1933; Weiskittel et al., 2009), and that the differences in slope are telling of different species’ life history patterns, as well as abiotic limitations. Specific species dominance was a significant factor, able to establish a distinct and constant slope of self- thinning. Previous studies of mixed-species stands were only able to account for populations of a limited geographic scope and specific species-composition ratios, which has relatively little context outside of their intended study area (Solomon and Zhang, 2002;

Stout and Nyland, 1986; Sturtevant et al., 1998). Providing regional estimates based on the estimation of plot species dominance, obviates the necessity for the construction of specific species ratios, and simplifies the estimation of potential SDImax. Individual plot estimates of

SDImax, achieved through a varying intercept, allowed for the both the assessment of individual stand’s self-thinning trajectory, which when pooled, contributed to the overall development of the self-thinning line, and individualized estimates of the SDImax., which are more reflective of a gradient of SDImax values across the region due to differences in abiotic conditions. Climate was found to be an important determinant in establishing patterns of

SDImax for each species across the landscape. This Information can be used to manage stands as climate changes.

4.1.2. To Characterize the Distribution and Abundance of the Important Species in the

Spruce-Fir Forest, while Comparing the Usefulness of Both Presence/Absence and

Abundance Models, as well as Alternatives, for Conservation Decisions

Presence/absence models were able to accurately predict and determine current distribution. Area under receiver operator curve (AUC) values for models averaged 0.99 ±

0.01 (mean ± SD), well above the accepted standard for excellent model performance, and almost errors were concentrated as false predictions of presence. Predicted presence for

119 each species exceeded known presence on the current landscape, as the models were able to determine locations that meet species climate requirements, while actual presence is limited by biotic factors, including competition, as well as further abiotic controls, include soil and nutrient availability.

Presence/absence models were found to predict with more accuracy than the abundance models, but this is not surprising considering the range of values is infinitely greater for abundance models than the binomial prediction for presence/absence models, and the fine spatial resolution used in this analysis. All abundance models underpredicted actual quantities, but were able to maintain relative patterns of abundance across the landscape. Of the three abundance dependent variables, basal area (BA; m2 ha-1), performed the best (Mean: 77.14 (±0.02)), while stem count (trees ha-1 (TPH)) performed the worst (Mean: 74.43 (±0.03)). The importance value (IV) performed slightly worse than the basal area models (Mean: 75.57 (±306.14), but benefits from being able reflect both BA and stem count, and this metric can be directly compared with previous, coarser resolution analyses for these species (Iverson et al. 2008).

The likelihood prediction object from the presence/absence models is able to reflect cores of abundance. Spearman’s indicated a strong positive relationship between all likelihood objects and BA abundance models. Average correlation was 0.90, with black spruce BA abundance exhibiting the strongest relationship (0.95) with the likelihood object, and red spruce the weakest (0.84). This is an important interpretation of the likelihood object, as it is generated from the computationally more efficient, and readily available,

120 presence/absence metric, while producing the same type of information: The core of abundance which is thought to represent habitat more optimal for the modeled species.

4.1.3. To Compare and Illustrate the Differences between the Results and Application of

Directly Calculated Variables Useful for Passive Management versus Predicted Variables

Useful for the Active Management of Forests

Overall, the SDImax metric correlated the best with climatic variables (83.65

(±11167.58)), when compared to alternative continuous variables. It is believed that since the SDImax was calculated only for plots dominated by either a spruce or fir species, that the dataset was more representative of optimal spruce or fir habitat. This resulted in a model that was able to better capture the climatic relationship, both in terms of pseudo R2 and, also, the selection of most important climatic variables. While the variables selected in

Chapter 2 were remarkably consistent between species, almost no overlap was shared between species in Chapter 3, representing the models’ ability to learn species specific climate signals. The overprediction of low values, seen throughout all random forest models built in this study, was particularly exacerbated in the SDImax metric models. It is believed that since SDImax is a stand level calculation where low numbers are inherently absent, that the model sought to exact a more normal distribution on the landscape.

SDImax models can be utilized for the construction of DMDs and the active management of future landscapes. While presence/absence models are important for understanding the full range of climatically suitable habitats, and abundance values provide the ability to prioritize suitable habitat based upon higher abundance, both of these are unable to assist forest managers in future forest planning. The predicted SDImax values for

121 each species, represent a regional species boundary line, or the maximum size-density relationship for a give species across a wide range of environments (Weller, 1990). The achievement of these predicted maximum values in a given stand is dependent upon the presence of ideal abiotic conditions. Climate was found to be partially accountable for this set of idyllic circumstances, and thus as climate changes, it is expected that forest stands that were previously limited to their population boundary lines, will be able achieve higher values found along the species boundary line.

Forest managers can use DMDs constructed from these models, both in the short- and long-term. In the short term, rapid migration of forest species is not expected in most environments, and forest composition may persist, though structure (Dolanc et al., 2013) and growth (Mohan et al., 2009) will likely change. As the climate transforms to conditions considered optimal for spruce and fir dominated stands, structure and composition can be managed for forest health enhancement, including increased resistance to the consequent effects of climate change. Spruce-fir forest types of the Acadian Region are naturally composed of multiple age classes and sizes, yielding various micro-environments for different species, and these inherent qualities are conducive to multi-age management.

Multi-age management has been lauded for its ability to increase forest resistance and resilience, as well as stand complexity and response diversity, by integrating partial disturbance into the management structure (D’Amato et al., 2011; O’Hara and Ramage,

2013). As climate changes, the risk of disturbance to forest ecosystems is expected to rise, if forest vulnerability is not reduced. Further benefits from irregular regeneration methods, or any management regime that reduce stand density, could include the reduction of competition for water and nutrients (Chmura et al., 2011) and decreased onset of mortality

122 due to drought (Elkin et al., 2015), and increased stand resistance to disease outbreak, as pest and diseases tend to be mono-specific (Edmonds et al., 2010). Alternatively, stands at low risk to fire and disease outbreak could be managed for climate change mitigation by increasing stocking levels, and therefore carbon storage, by retaining older mature individuals present in the canopy (D’Amato et al., 2011; Nunery and Keeton, 2010).

On a longer time scale, species composition and structure, particularly in regards to density, will continually need to be managed, but species will likely shift to new habitats.

The future suitable habitat predictions provided by the presence/absence and abundance models, can assist in determining ideal locations for future habitat and conservation prioritization. Novel stand species compositions will likely appear during this time, as well as corresponding interactions in regards to interspecific competing life history strategies and stand development (Williams et al., 2013). While the models presented in this thesis do not directly account for biotic interactions, SDImax models similar to the ones constructed here will continue to be useful, as the only requirement for utilization is a specific species dominance. However, SDImax models will need to be reconfigured with datasets from future stands, as life history strategies, and the corresponding realization on the landscape, are thought to be an important factor in determining species specific self-thinning line coefficients. In regards to the migration of species to future suitable habitats, shifts in habitat are likely to outpace many species’ ability to disperse and migrate. Species with large distributions that cover numerous environmental gradients, are likely phenologically predisposed to adaptation to shifts in climate in their current range (Aitken et al., 2008).

Species with small, specific ranges, which are centered in abundance at the core, and species with low fecundity, are at risk for extirpation. These at-risk species might rely on

123 locations of refugia, which are locations of limited spatial extent that are environmentally suitable for species to retract to during times of climate stress (Keppel et al., 2012).

Protecting locations for refugia, as well as the establishment of populations through facilitated migration outside of current habitat (i.e. ex situ refugia), might assist species with heightened longevity, large phenotypic plasticity, and low dispersal, as these nucleated populations might persist until the eventual arrival of additional member of the species, or the return of suitable climate. The SDM outputs generated in this thesis, along with abiotic overlays and mechanistic model outputs, can help determine locations of suitable refugia.

4.2 .Summary of Findings by Species

Previous studies have found that variation in model performance is greater among tree species than among techniques (Guisan et al., 2007), and that no technique can rescue species that are difficult to predict. This thesis confirmed this trend, with consistent ranking in model performance amongst the species analyzed here. Discussed below is model success amongst species, as well as the implication of specific life history requirements on the models, and how the models should be interpreted in regards to future distribution in management.

4.2.1. Balsam Fir (Abies balsamea L.)

Balsam fir responded well to all models presented in this thesis, performing best in response to the SDImax dependent variable. Balsam fir is a generalist with the ability to survive in a wide array of climate and soil conditions. This species is also extremely competitive and flowers and thrives in full light (Bakuzis and Hansen, 1965), and has increased in distribution across the landscape due to anthropogenic forest disturbance.

124

Models were thus able to establish and detect a larger share of this species fundamental niche in regards to climate, but realized habitat on non-disturbed landscapes is much more limited by biotic competition. Similarly, a predicted steeper slope of the self-thinning lines appears related to high stem density and lower average size values, which is currently associated with balsam fir due to alterations to the natural disturbance regime in this region. SDImax values and the associated self-thinning line will need to be reconfigured depending on future forest development and disturbance. Due to this species’ comparative tolerance for warmer temperatures, large range, high abundance, high fecundity, and competitive superiority in disturbed environments, the future outlook for suitable realized habitat under climate change is positive.

4.2.2. White Spruce (Picea glauca (Moench) Voss)

White spruce consistently performed the worst amongst the tree species considered in this analysis, with the lowest reported AUC and psuedo R2 values. Of the continuous variables, white spruce responded the best to SDImax modeling. Lower accuracy amongst models is likely a result of this species sparse and inconsistent distribution across the landscape, due to more exacting light and soil conditions than associated conifers

(Kabzems, 1971). Currently, in northeastern North America, white spruce grows abundantly in the moist cool fog belt of the Acadian coast, and it is found in the interior of Maine and elsewhere in low abundance in mixed stands. White spruce has increased in dominance across the landscape due to its ability to thrive in farm fields abandoned over the last century, though it is outcompeted over time (Mosseler et al., 2003). Both current and future suitable habitat based on important climate variables predicted large areas of suitable habitat for this species, while actual realized habitat is much more restricted. White spruce

125 is considered a plastic species and is able to grow in a variety of climatic conditions.

Generally, plastic species’ ranges are larger than those with more specific niches (Morin and

Lechowicz, 2013), and abundance and frequency of these species within their range are controlled less by abiotic factors, and more by biotic competition (Murphy et al., 2006).

Pollen records show that this species rapidly established on the post-glaciated landscape at the end of the last glacial maximum. The success of white spruce at this time is due to lack of biotic completion, and the presence of rich, coarse-textured soils with good drainage

(Lindbladh et al., 2007), that quickly disappeared as the climate became colder and wetter

(Grimm and Jacobson, 2003). While white spruce likely has the phenotypic plasticity to be able to survive in a variety of climate conditions (Gordon, 1996), it suffers from the ability to adequately migrate due to its restricted current range. Though the possibility exists that white spruce may be able to compete as other species concurrently decline in fitness due to climate change, this species is a good candidate for current habitat protection and facilitated migration through the establishment of ex situ refugia in proper suitable habitat in northeastern Canada.

4.2.3. Black Spruce (Picea mariana (Miller) B.S.P.)

Black spruce consistently performed the best amongst the tree species considered in this analysis in regards to the abundance models. Black spruce was the only species whose model performance did not increase with the use of SDImax as the dependent variable. The R2 values for this model were still high, but the random forest algorithm likely had an easier time detecting the consistent and expansive abundance metrics across the landscape. Black spruce, also considered a plastic species, can survive in a wide variety of climatic conditions. Studies indicate that this species experienced large range shifts in a

126 relatively quick period at the end of the last ice age, and that there is no apparent sign that dispersal limitation is constraining their modern range (Williams et al., 2013). Location of black-spruce dominated krummholz at the tree line, as well as habitat at the forest-tundra boundary, are expected to continue to react positively to climate warming (Gamache and

Payette, 2004; Thomson et al., 2009). Black spruce will likely continue to form uniform, high stem-density, low individual tree-size stands on poor sites in northern Canada, though the locations of these habitats will shift north. Density management should be considered when feasible to reduce the risk of fire in these regions. In the Acadian Region, black spruce habitat may persist in krummholz and other high elevation locations. The peatlands of

Maine and other moist habitats, currently suitable habitat for black spruce in the Acadian

Region, may persist if current hydrology is maintained (Anderson and Davis, 1997) and if other wet-footed species, such as tupelo (Nyssa sylvatica), do not prohibitively compete for resources in this environment as they migrate north.

4.2.4. Red Spruce (Picea rubens Sarg.)

Red spruce responded well to all models presented in this thesis, due to an easy to detect, small and specific range. It responded the best amongst models and species to the

SDImax dependent variable. Red spruce dominated plots captured in the SDImax analysis, primarily represent mature stands with red spruce in the overstory, and this model was able to detect the suite of climatic variables that support this habitat. Red spruce is a temperate species and its habitat is limited by sensitivity to low winter temperatures (Thompson et al.,

2006), and reliant upon adequate moisture in cool environs (Dumais and Prévost, 2007).

Models indicate that suitable habitat will persist in the Acadian Region, particularly when compared to its fellow Picea species. While it is unclear where red spruce endured during

127 the late-glacial and early Holocene, it was a late arrival to northeastern North America, initially limited by cold temperatures during the Younger Dryas, and later prohibited by seasonal temperature extremes, including dry and warm fire-prone summer conditions, caused by solar precession (Grimm and Jacobson, 2003; Lindbladh et al., 2003). Climate predictions for the Acadian Region do predict increases in summer climate extremes, but winter temperatures are also expected to warm, and future suitable habitat for this species is predicted throughout northeastern North America. Warmer temperatures will increase growth in red spruce habitat that is currently surviving at the edge of its cold tolerance, such as krummholz and other Acadian high altitude environments. Migratory success of red spruce is uncertain, as the species is extremely shade tolerant and establishment is best on non-disturbed landscapes and in the understory of pre-existing stands. Due to the longevity of the species and its preference of warmer temperatures, red spruce is a good candidate for facilitated migration and the establishment of ex situ refugia, though providing the necessary conditions for establishment need to be carefully considered. The relative rarity of this species across the landscape, necessitates that current remaining old growth habitat, including high altitude elevations, be preserved.

4.3. Conclusion

This thesis set out to link species specific data with climate and topographic variables in order to generate models that would accurately predict changes in species distribution due to climate change. While the analyses presented here were able to characterize species’ known or realized distribution with climate with high statistical significance, many questions are left unanswered on what future forests will look like and how we will get there. In regards to species’ climatic tolerances, only known distributions,

128 which included historical inventories, were analyzed, and it is likely that species’ fundamental niches are much wider than what we see on the current landscape.

Understanding species’ full range of climatic tolerance in order to predict future suitable habitat requires more knowledge about their fundamental niche, as well as the genetic expression of tolerance to different climate variables. Further research, including paleoecological climate research and provenance testing of important species, needs to be under taken in order to understand to better predict relationships.

Even if species-climate associations were fully characterized in this study, a whole suite of other factors that determine species occurrence and dominance on the landscape were not captured in these analyses. Perhaps most importantly, it is very difficult to predict future biotic interactions and the effects on forest structure and function. Dynamic landscapes predictions can be achieved with mechanistic models, and while computer capacity currently limits region-wide studies, even pocket analyses help elucidate future interactions between species on the landscape. The real limit to these mechanistic models though is our knowledge of how species’ will redistribute and disperse in novel climate and competitive environments, as most of our data is based on the present observed world.

Currently, effects of climate change to forest ecosystems, such as tree-line advances and shifts to higher latitudes or elevations, are not occurring at the rate that would be expected given the change in climate, and in fact, the opposite reaction has been observed in some locations. Many studies have focused on future predictions to ecosystems due to climate change, but more research needs to be undertaken that observes the current effects of climate to forest ecosystems, and what the drivers of these changes are.

129

BIBLIOGRAPHY

Adams, H.D., Guardiola-Claramonte, M., Barron-Gafford, G.A., Villegas, J.C., Breshears, D.D., Zou, C.B., Troch, P.A., Huxman, T.E., 2009. Temperature sensitivity of drought-induced tree mortality portends increased regional die-off under global-change-type drought. Proc. Natl. Acad. Sci. U. S. A. 106, 7063–7066. doi:10.1073/pnas.0901438106

Aitken, S.N., Yeaman, S., Holliday, J.A., Wang, T., Curtis-McLane, S., 2008. Adaptation, migration or extirpation: climate change outcomes for tree populations. Evol. Appl. 1, 95–111. doi:10.1111/j.1752-4571.2007.00013.x

Allen, C.D., Macalady, A.K., Chenchouni, H., Bachelet, D., McDowell, N., Vennetier, M., Kitzberger, T., Rigling, A., Breshears, D.D., Hogg, E.H. (Ted), Gonzalez, P., Fensham, R., Zhang, Z., Castro, J., Demidova, N., Lim, J.H., Allard, G., Running, S.W., Semerci, A., Cobb, N., 2010. A global overview of drought and heat-induced tree mortality reveals emerging climate change risks for forests. For. Ecol. Manage. 259, 660–684. doi:10.1016/j.foreco.2009.09.001

Anderson, B.T., Hayhoe, K., Liang, X.-Z., 2010. Anthropogenic-induced changes in twenty- first century summertime hydroclimatology of the Northeastern US. Clim. Change 99, 403–423. doi:10.1007/s10584-00-9674-3

Anderson, D.S., Davis, R.B., 1997. The vegetation and its environments in Maine peatlands. Can. J. Bot. 75, 1785–1805. doi:10.1139/b97-893

Araújo, M., Pearson, R., Thuillers, W., Erhard, M., 2005. Validation of species–climate impact models under climate change. Glob. Chang. Biol. 11, 1504–1513.

Atlantic Provinces Economic Council, 2003. The New Brunswick forest industry: The potential economic impact of proposal to increase the wood supply. Halifax, NS.

Atlantic Provinces Economic Council, 2005. The economic impact of the Nova Scotia forest industry. Halifax, NS.

Austin, M., 2007. Species distribution models and ecological theory: A critical assessment and some possible new approaches. Ecol. Modell. 200, 1–19. doi:10.1016/j.ecolmodel.2006.07.005

Bakuzis, E. V., Hansen, H.L., 1965. Balsam fir-a monographic review. University of Minnesota Press, Minneapolis, MN.

Barry, S.C., Welsh, A.H., 2002. Generalized additive modelling and zero inflated count data. Ecol. Modell. 157, 179–188.

130

Barton, A., White, A., Cogbill, C., 2012. The changing nature of the Maine woods. University Press of New England, Durham, NH, USA.

Bechtold, W.A., Patterson, P.L. (Eds.), 2005. The enhanced forest inventory and analysis program - National sampling design and estimation procedures. USDA For. Ser. Gen. Tech. Rep. SRS-80.

Beckage, B., Osborne, B., Gavin, D.G., Pucko, C., Siccama, T., Perkins, T., 2008. A rapid upward shift of a forest ecotone during 40 years of warming in the Green Mountains of Vermont. Proc. Natl. Acad. Sci. U. S. A. 105, 4197–4202. doi:10/1073/pnas.0708921105

Beers, T.W., Dress, P.E., Wensel, L.C., 1966. Notes and observations: Aspect transformation in site productivity research. J. For. 64, 691–692.

Bi, H., 2004. Stochastic frontier analysis of a classic self-thinning experiment. Austral Ecol. 29, 408–417. doi:10.1111/j.1442-9993.2004.01379.x

Bradbury, J., Keim, B., Wake, C., 2002. US East Coast trough indices at 500 hPa and New England winter climate variability. J. Clim. 15, 3509–3518.

Breiman, L., 2001. Random Forests. Mach. Learn. 45, 5–32.

Brenning, A., 2008. Statistical geocomputing combining R and SAGA: The example of landslide susceptibility analysis with generalized additive models, in: Boehner, J. Blaschke, T. Montanarella, L. (Ed.), SAGA - Seconds Out (Hamburger Beitraege Zur Physischen Geographie Und Landschaftsoekologie, Vol. 19). pp. 23–32.

Brumelis, G., Carleton, T.J., 1988. The vegetation of postlogged black spruce lowlands in central Canada. I. Trees and tall shrubs. Can. J. For. Res. 18, 1470–1478. doi:10.1139/x88-226

Bush, E.J., Loder, J.W., James, T.S., Mortsch, L.D., Cohen, S.J., 2014. Chapter 2: An overview of Canada’s changing climate, in: Warren, F.J., Lemmen, D.S. (Eds.), Canada in a Changing Climate: Sector Perspectives on Impacts and Adaptations. Government of Canada, Ottawa, ON, pp. 23–64.

Cade, B.S., Guo, Q., 2000. Estimating effects of constraints on plant performance with regression quantiles. Oikos 91, 245–254. doi:10.1034/j.1600-0706.2000.910205.x

Cade, B.S., Noon, B.R., 2003. A gentle introduction to quantile regression for ecologists in a nutshell. Front. Ecol. Environ.

Cade, B.S., Terrell, J.W., Schroeder, R.L., 1999. Estimating effects of limiting factors with regression quantiles. Ecology 80, 311–323.

131

Canada’s National Forest Inventory, 2006. Total tree volume (million m3) on forest land by forest type, age class, and terrestrial ecozone in Canada.

Cao, Q.V., Dean, T.J., 2015. Using nonlinear quantile regression to estimate the self-thinning boundary curve, in: Holley, A.G., Connor, K.F., Haywood, J.D. (Eds.), Proceedings of the 17th Biennial Southern Silvicultural Research Conference. E-Gen. Tech. Rep. SRS-203. U.S. Department of Agriculture, Forest Service, Southern Research Station, Asheville, N.C., pp. 349–351.

Chen, C., Liaw, A., Breiman, L., 2004. Using random forest to learn imbalanced data. Univ. California, Berkeley 1–12.

Chen, I.-C., Hill, J.K., Ohlemüller, R., Roy, D.B., Thomas, C.D., 2011. Rapid range shifts of species associated with high levels of climate warming. Science 333, 1024–1026. doi:10.1126/science.1206432

Chmura, D.J., Anderson, P.D., Howe, G.T., Harrington, C. a., Halofsky, J.E., Peterson, D.L., Shaw, D.C., Brad St.Clair, J., 2011. Forest responses to climate change in the northwestern United States: Ecophysiological foundations for adaptive management. For. Ecol. Manage. 261, 1121–1142. doi:10.1016/j.foreco.2010.12.040

Chok, N.S., 2010. Pearson’s versus Spearman's and Kendall's correlation coefficients for continuous data. M.S. Thesis. University of Pittsburgh.

Cogbill, C., 2000. Vegetation of the presettlement forests of northern New England and New York. Rhodora 102, 250–276.

Cogbill, C., Burk, J., Motzkin, G., 2002. The forests of presettlement New England, USA: spatial and compositional patterns based on town proprietor surveys. J. Biogeogr. 29, 1279–1304.

Cogbill, C., White, P., Wiser, S., 1997. Predicting treeline elevation in the southern Appalachians. Castanea 62, 137–146.

Cohen, J., 1960. A coefficient of agreement for nominal scales. Educ. Psychol. Meas. 20, 37– 46.

Condit, R., Hubbell, S.P., Foster, R.B., 1996. Changes in tree species abundance in a Neotropical forest: impact of climate change. J. Trop. Ecol. 12, 231–256. doi:10.1017/S0266467400009433

Curtis, J.T., McIntosh, R.P., 1951. An upland forest continuum in the prairie-forest border region of Wisconsin. Ecology 32, 476–496. doi:10.2307/1931725

132

Cutler, D.R., Edwards, T.C., Beard, K.H., Cutler, A., Hess, K.T., Gibson, J., Lawler, J.J., 2007. Random forests for classification in ecology. Ecology 88, 2783–92.

D’Amato, A.W., Bradford, J.B., Fraver, S., Palik, B.J., 2011. Forest management for mitigation and adaptation to climate change: Insights from long-term silviculture experiments. For. Ecol. Manage. 262, 803–816. doi:10.1016/j.foreco.2011.05.014

Davis, M.B., 1976. The Pleistocene biogeography of temperate decidous forests, in: Ecology of the Pleistocene: A Symposium. p. 13.

Davis, R., 1966. Spruce-fir forests of the coast of Maine. Ecol. Monogr. 36, 79–94.

Davis, W.C., 1991. The role of advanced growth in regeneration of red spruce and balsam fir in east central Maine, in: Simpson, C.M. (Ed.), Proceedings of the Conference on Natural Regeneration Management. Forestry Canada, Maritimes Region, Fredericton, NB, pp. 157–168.

Dean, T.J., Long, J.N., 1992. Influence of leaf area and canopy structure on size-density relations in even-aged lodgepole pine stands. For. Ecol. Manage. 49, 109–117. doi:10.1016/0378-1127(92)90164-5

DeBell, D.S., Harms, W.R., Whitesell, C.D., 1989. Stockability: A major factor in productivity differences between pinus taeda plantations in Hawaii and the Southeastern United States. For. Sci. 35, 708–719.

Deluca, W. V., King, D.I., 2014. Influence of hiking trails on montane birds. J. Wildl. Manage. 78, 494–502.

Dolanc, C.R., Thorne, J.H., Safford, H.D., 2013. Widespread shifts in the demographic structure of subalpine forests in the Sierra Nevada, California, 1934 to 2007. Glob. Ecol. Biogeogr. 22, 264–276. doi:10.1111/j.1466-8238.2011.00748.x

Doyle, T.W., Krauss, K.W., Conner, W.H., From, A.S., 2010. Predicting the retreat and migration of tidal forests along the northern Gulf of Mexico under sea-level rise. For. Ecol. Manage. 259, 770–777. doi:10.1016/j.foreco.2009.10.023

Drew, T., Flewelling, J.W., 1977. Some recent Japanese theories of yield-density relationships and their application to Monterey pine plantations. For. Sci. 23, 517–534.

Ducey, M.J., Knapp, R.A., 2010. A stand density index for complex mixed species forests in the northeastern United States. For. Ecol. Manage. 260, 1613–1622. doi:10.1016/j.foreco.2010.08.014

133

Dumais, D., Prévost, M., 2007. Management for red spruce conservation in Quebéc: The importance of some physiological and ecological characteristics-A review. For. Chron. 83, 378–392.

Duveneck, M., Scheller, R., White, M., 2014. Climate change effects on northern Great Lake (USA) forests: A case for preserving diversity. Ecosphere 5, 1–26.

Edmonds, R.L., Agee, J.K., Gara, R.I., 2010. Forest Health and Protection. Waveland Press, Inc., Long Grove, IL.

Elith, J., Graham, C.H., Anderson, R.P., Dudi, M., Ferrier, S., Guisan, A., Hijmans, R., Hettmann, F., Leathwik, J.R., Lehmann, A., Li, J., Lohmann, L.G., B.A., L., Manion, G., Mortiz, C., Nakamura, M., Nakazawa, Y., Overton, J.M., Townsend Peterson, A., Phillips, S.J., Richardson, K., Scachetti-Pereira, R.S., Schapire, R.E., Soberon, J., Williams, S., Wisz, M.S., N.E., Z., 2006. Novel methods improve prediction of species’ distributions from occurrence data. Ecography 29, 129–151.

Elith, J., Kearney, M., Phillips, S., 2010. The art of modelling range-shifting species. Methods Ecol. Evol. 1, 330–342. doi:10.1111/j.2041-210X.2010.00036.x

Elith, J., Leathwick, J.R., 2009. Species distribution models: Ecological explanation and prediction across space and time. Annu. Rev. Ecol. Evol. Syst. 40, 677–697. doi:10.1146/annurev.ecolsys.110308.120159

Elkin, C., Guiggiola, A., Rigling, A., Bugmann, H., 2015. Short- and long-term efficacy of forest thinning to mitigate drought impacts in mountain forests in the European Alps. Ecol. Appl. 25, 1083–1098.

Eyre, F.H., 1980. Forest cover types of the United States and Canada.

Fielding, A.H., Bell, J.F., 1997. A review of methods for the assessment of prediction errors in conservation presence/absence models. Environ. Conserv. 24, 38–49. doi:10.1017/S0376892997000088

Forest2Market, 2009. The economic impact of privately-owned forests, in: Report for National Alliance of Forest Owners. Charlotte, NC, p. 19.

Foster, D.R., 1983. The history and pattern of fire in the boreal forest of southeastern Labrador. Can. J. Bot. 61, 2459–2471. doi:10.1139/b83-269

Frank, R.M., Bjorkbom, J.C., 1973. A silvicultural guide for spruce-fir in the Northeast.

Fraver, S., Seymour, R.S., Speer, J.H., White, A.S., 2007. Dendrochronological reconstruction of spruce budworm outbreaks in northern Maine, USA. Can. J. For. Res. 37, 523–529.

134

Gajewski, K., 1988. Late holocene climate changes in eastern North America estimated from pollen data. Quat. Res. 29, 255–262.

Gamache, I., Payette, S., 2004. Height growth response of tree line black spruce to recent climate warming across the forest-tundra of eastern Canada. J. Ecol. 92, 835–845.

Gawler, S., Cutko, A., 2010. Natural Landscapes of Maine: A guide to Natural Communites and Ecosystem. Maine Natural Areas Program, Maine Department of Conservation, Augusta, Maine.

Gelman, A., Hill, J., 2006. Data Analysis Using Regression and Multilevel/Hierarchical Models. Cambridge University Press, Cambridge, United Kingdom and New York, NY, USA.

Geraci, M., 2014. Linear quantile mixed models: the lqmm package for Laplace quantile regression. J. Stat. Softw. 57, 1–29.

Geraci, M., Bottai, M., 2007. Quantile regression for longitudinal data using the asymmetric Laplace distribution. Biostatistics 8, 140–154. doi:10.1093/biostatistics/kxj039

Gillett, N.P., Weaver, A.J., Zwiers, F.W., Flannigan, M.D., 2004. Detecting the effect of climate change on Canadian forest fires. Geophys. Res. Lett. 31. doi:10.1029/2004GL020876

Gordon, A., 1996. The sweep of the boreal in time and space, from forest formations to genes, and implications for management. For. Ch 72.

Gordon, A.G., 1976. The taxonomy and genetics of Picea rubens and its relationship to Picea mariana. Can. J. Bot. 54, 781–813.

Graham, C.H., Elith, J., Hijmans, R.J., Guisan, A., Townsend Peterson, A., Loiselle, B.A., Group, T.N.P.S.D.W., 2007. The influence of spatial errors in species occurrence data used in distribution models. J. Appl. Ecol. 45, 239–247. doi:10.1111/j.1365- 2664.2007.01408.x

Greenwood, M.S., O’Brien, C.L., Schatz, J.D., Diggins, C.A., Day, M.E., Jacobson, G.L., White, A.S., Wagner, R.G., 2008. Is early life cycle success a determinant of the abundance of red spruce and balsam fir? Can. J. For. Res. 38, 2295–2305. doi:10.1139/X08-072

Grimm, E., Jacobson, G., 2003. Late-Quaternary vegetation history of the eastern United States. Dev. Quat. Sci. 1, 381–402. doi:10.1016/S1571-0866(03)01017-0

Guisan, A., Tingley, R., Baumgartner, J.B., Naujokaitis-Lewis, I., Sutcliffe, P.R., Tulloch, A.I.T., Regan, T.J., Brotons, L., McDonald-Madden, E., Mantyka-Pringle, C., Martin, T.G., Rhodes, J.R., Maggini, R., Setterfield, S. a, Elith, J., Schwartz, M.W., Wintle, B. a,

135

Broennimann, O., Austin, M., Ferrier, S., Kearney, M.R., Possingham, H.P., Buckley, Y.M., 2013. Predicting species distributions for conservation decisions. Ecol. Lett. `6, 1424–1435. doi:10.1111/ele.12189

Guisan, A., Zimmermann, N.E., Elith, J., Graham, C.H., Phillips, S.P., Peterson, A.T., 2007. What matters for predicting the occurrences of trees: techniques, data, or species characteristics? Ecol. Monogr. 77, 615–630.

Halliday, W.. E.. D.., Brown, A.. . . A., 1943. The Distribution of Some Important Forest Trees in Canada. Ecology 24, 353–373.

Hanberry, B.B., Palik, B.J., He, H.S., 2012. Comparison of historical and current forest surveys for detection of homogenization and mesophication of Minnesota forests. Landsc. Ecol. 27, 1495–1512. doi:10.1007/s10980-012-9805-5

Hann, D.W., 2014. Modeling of the maximum size-density line and its trajectory line for tree species: Observations and opinions. For. Biometrics Res. Pap. 5 1–33.

Hansen, A., Neilson, R., Dale, V., Flather, C., Iverson, L., Currie, D., Shafer, S., Cook, R., Bartlein, P., 2001. Global change in forests: responses of species, communities, and biomes. Bioscience 51, 765–779.

Harms, W.R., DeBell, D.S., Whitesell, C.D., 1994. Stand and tree characteristics and stockability in Pinus taeda plantations in Hawaii and South Carolina. Can. J. For. Res. 24, 511–521.

Harsch, M.A., Hulme, P.E., McGlone, M.S., Duncan, R.P., 2009. Are treelines advancing? A global meta-analysis of treeline response to climate warming. Ecol. Lett. 12, 1040– 1049. doi:10.1111/j.1461-0248.2009.01355.x

Hart, A.C., 1965. Silvical characteristics of red spruce (Picea rubens), in: USDA Forest Service, Station Paper 124. Northeastern Forest Experiment Station, Upper Darby, PA, p. 18.

Hayhoe, K., Wake, C., Anderson, B., Liang, X.-Z., Maurer, E., Zhu, J., Bradbury, J., DeGaetano, A., Stoner, A.M., Wuebbles, D., 2008. Regional climate change projections for the Northeast USA. Mitig. Adapation Strateg. Glob. Chang. 13, 425–436. doi:10.1007/s11027-007-9113-2

Hayhoe, K., Wake, C.P., Huntington, T.G., Luo, L., Schwartz, M.D., Sheffield, J., Wood, E., Anderson, B., Bradbury, J., DeGaetano, A., Troy, T.J., Wolfe, D., 2007. Past and future changes in climate and hydrological indicators in the US Northeast. Clim. Dyn. 28, 381– 407. doi:10.1007/s00382-006-0187-8

136

Heikkinen, R., Luoto, M., Araújo, M., Virkkala, R., Thuillers, W., Sykes, M., 2006. Methods and uncertainties in bioclimatic envelope modelling under climate change. Prog. Phys. Geogr. 6, 751–777.

Hijmans, R.J., 2014. Raster: Geographic data analysis and modeling. R package version 2.2- 12.

Hodgkins, G.A., James Ivan, I.C., Huntington, T.G., 2002. Historical changes in lake ice-out dates as indicators of climate change in New England, 1850-2000. Int. J. Climatol. 22, 1819–1827. doi:10.1002/joc.857

Hosmer, R., 1902. A study of the Maine spruce, Maine State Forest Commissioner Annual Report. Augusta, Maine.

Huntington, T.G., Richardson, A.D., McGuire, K.J., Hayhoe, K., 2009. Climate and hydrological changes in the northeastern United States: recent trends and implications for forested and aquatic ecosystems. Can. J. For. Res. 39, 199–212. doi:10.1139/X08-116

Huntley, B., 1991. How plants respond to climate change: Migration rates, individualism and the consequences for plant communities. Ann. Bot. 67, 15–22.

Iverson, L.R., Prasad, A.M., 1998. Predicting abundance of 80 tree species following climate change in the eastern United States. Ecol. Monogr. 68, 465. doi:10.2307/2657150

Iverson, L.R., Prasad, A.M., Matthews, S.N., Peters, M., 2008. Estimating potential habitat for 134 eastern US tree species under six climate scenarios. For. Ecol. Manage. 254, 390–406. doi:10.1016/j.foreco.2007.07.023

Iverson, L.R., Prasad, A.M., Matthews, S.N., Peters, M.P., 2010. Merger of three modeling approaches to assess potential effects of climate change on trees in the eastern United States, in: Azevedo, J.C., Feliciano, M., Castro, J., Pinto, M.A. (Eds.), Forest Landscapes and Global Change. New Frontiers in Management, Conservation, and Restoration. Braganca, Portugal, pp. 135–141.

Iverson, L.R., Prasad, A.M., Matthews, S.N., Peters, M.P., 2011. Lessons learned while integrating habitat, dispersal, disturbance, and life-history traits into species habitat models under climate change. Ecosystems 14, 1005–1020. doi:10.1007/s10021-011- 9456-4

Jack, S.B., Long, J.N., 1996. Linkages between silviculture and ecology: An analysis of density management diagrams. For. Ecol. Manage. 86, 205–220. doi:10.1016/S0378- 1127(96)03770-X

Jacobson, G.L., Fernandez, I.J., Mayewski, P.A., Schmitt, C. V (Eds.), 2009. Maine’s Climate Future: An Initial Assessment. Univeristy of Maine, Orono, ME.

137

Jones, B.S., 2007. Multilevel Models.

Joyce, D.G., Rehfeldt, G.E., 2013. Climatic niche, ecological genetics, and impact of climate change on eastern white pine (Pinus strobus L.): Guidelines for land managers. For. Ecol. Manage. 295, 173–192. doi:10.1016/j.foreco.2012.12.024

Jung, S., 1996. Quasi-likelihood for median regression models. J. Am. Stat. Assoc. 91, 251– 257.

Kabzems, A., 1971. The growth and yield of well stocked white spruce stands in the mixedwood section of Saskatchewan. Prince Albert, SA.

Kearney, M., Porter, W., 2009. Mechanistic niche modelling: Combining physiological and spatial data to predict species’ ranges. Ecol. Lett. 12, 334–350. doi:10.1111/j.1461- 0248.2008.01277.x

Keim, B.D., 1998. A climate primer for New England, in: New England’s Changing Climate Weather and Air Quality. University of New Hampshire Climate Change Research Center, Durham, NH, USA.

Keith, D.A., Akçakaya, H.R., Thuiller, W., Midgley, G.F., Pearson, R.G., Phillips, S.J., Regan, H.M., Araújo, M.B., Rebelo, T.G., 2008. Predicting extinction risks under climate change: coupling stochastic population models with dynamic bioclimatic habitat models. Biol. Lett. 4, 560–563. doi:10.1098/rsbl.2008.0049

Kelly, A.E., Goulden, M.L., 2008. Rapid shifts in plant distribution with recent climate change. Proc. Natl. Acad. Sci. U. S. A. 105, 11823–11826. doi:10.1073/pnas.0802891105

Kelty, M.J., D’Amato, A.W., 2006. A historical perspective on diameter-limit cutting in the Northeast., in: Kenefic, L.S., Nyland, R.D. (Eds.), Proceedings of the Conference on Diameter-Limit Cutting in Northeastern Forests. USDA Forest Service. Gen. Tech. Rep. NE- 342.

Kent, M., Coker, P., 1992. Vegetation description and data analysis: a practical approach. John Wiley and Sons, Chichester, West Sussex, UK.

Keppel, G., Van Niel, K.P., Wardell-Johnson, G.W., Yates, C.J., Byrne, M., Mucina, L., Schut, A.G.T., Hopper, S.D., Franklin, S.E., 2012. Refugia: Identifying and understanding safe havens for biodiversity under climate change. Glob. Ecol. Biogeogr. 21, 393–404. doi:10.1111/j.1466-8238.2011.00686.x

Klemm, O., Bachmeier, A.S., Talbot, R.W., Klemm, K.I., 1994. Fog chemistry at the New England coast: Influence of air mass history. Atmos. Environ. 28, 1181–1188.

138

Koenker, R., 2004. Quantile regression for longitudinal data. J. Multivar. Anal. 91, 74–89. doi:10.1016/j.jmva.2004.05.006

Koenker, R., Bassett Jr., G., 1978. Regression Quantiles. Econometrica 46, 33–50.

Koenker, R., Machado, J.A.F., 1999. Goodness of Fit and Related Inference Processes for Quantile Regression. J. Am. Stat. Assoc. 94, 1–22.

Kyker-Snowman, T., Clark, D., Eck, H., French, J., Morin, D., Stone, R., Fluet, M., Campbell, S., Small, D., Lyons, P., Zimmerman, J., 2007. Quabbin Reservoir watershed system: Land management plan. 2007–2017. MA Department of Conservation and Recreation, Division of Water Supply Protection, Office of Watershed Management.

Landsberg, J., 2003. Physiology in forest models: history and the future. For. Biometry Model. Inf. Sci. 1, 49–63.

Larocque, G.R., Bhatti, J.S., Boutin, R., Chertov, O., 2008. Uncertainty analysis in carbon cycle models of forest ecosystems: Research needs and development of a theoretical framework to estimate error propagation. Ecol. Modell. 219, 400–412. doi:10.1016/j.ecolmodel.2008.07.024

Lawless, J., 1987. Negative binomial and mixed Poisson regression. Can. J. Stat. 15, 209–225.

Leak, W.B., 1982. Habitat mapping and interpretation in New England. USDA For. Serv. Res. Pap. NE 496, 32.

Lenoir, J., Gégout, J.C., Marquet, P.A., de Ruffray, P., Brisse, H., 2008. A significant upward shift in plant species optimum elevation during the 20th century. Science 320, 1768– 1771. doi:10.1126/science.1156831

Li, R., Weiskittel, A.R., Kershaw, J.A., 2011. Modeling annualized occurrence , frequency , and composition of ingrowth using mixed-effects zero- inflated models and permanent plots in the Acadian Forest Region of North America. Can. J. For. Res. 41, 2077–2089. doi:10.1139/X11-117

Liaw, A., Wiener, M., 2002. Classification and regression by randomForest. R News 2/3, 18– 22.

Lima-Ribeiro, M.S., Hortal, J., Varela, S., Diniz-Filho, J.A.F., 2014. Constraint envelope analyses of macroecological patterns reveal climatic effects on Pleistocene mammal extinctions. Quat. Res. 82, 260–269. doi:http://dx.doi.org/10.1016/j.yqres.2014.02.003

Lindbladh, M., Jacobson, G.L., Schauffler, M., 2003. The postglacial history of three Picea species in New England, USA. Quat. Res. 59, 61–69. doi:10.1016/S0033- 5894(02)00023-6

139

Lindbladh, M., Oswald, W.W., Foster, D.R., Faison, E.K., Hou, J., Huang, Y., 2007. A late- glacial transition from Picea glauca to Picea mariana in southern New England. Quat. Res. 67, 502–508. doi:10.1016/j.yqres.2007.01.010

Little Jr., E.L., 1971. Atlas of United States trees. Volume 1. Conifers and important hardwoods. US Department of Agriculture, Forest Service, Washington, DC.

Logan, J.A., Régnière, J., Powell, J.A., 2003. Assessing the impacts of global warming on forest pest dynamics. Front. Ecol. Environ. 1, 130–137. doi:10.1890/1540- 9295(2003)001[0130:ATIOGW]2.0.CO;2

Long, J.N., 1985. A practical approach to density management. For. Chron. 61, 23–27. doi:doi:10.5558/tfc61023-1

Loo, J., Ives, N., 2003. The Acadian forest: Historical condition and human impacts. For. Chron. 79, 462–474.

Luoto, M., Pöyry, J., Heikkinen, R.K., Saarinen, K., 2005. Uncertainty of bioclimate envelope models based on the geographical distribution of species. Glob. Ecol. Biogeogr. 14, 575–584. doi:10.1111/j.1466-822x.2005.00186.x

Maini, J.S., 1966. Phytoecological study of sylvotundra at small tree lake, N.W.T. Arctic 19, 220–243.

Maiorano, L., Cheddadi, R., Zimmermann, N.E., Pellissier, L., Petitpierre, B., Pottier, J., Laborde, H., Hurdu, B.I., Pearman, P.B., Psomas, A., Singarayer, J.S., Broennimann, O., Vittoz, P., Dubuis, A., Edwards, M.E., Binney, H.A., Guisan, A., 2013. Building the niche through time: Using 13,000 years of data to predict the effects of climate change on three tree species in Europe. Glob. Ecol. Biogeogr. 22, 302–317. doi:10.1111/j.1466- 8238.2012.00767.x

Major, J.E., Barsi, D.C., Mosseler, A., Campbell, M., Rajora, O.P., 2003. Light-energy processing and freezing-tolerance traits in red spruce and black spruce: species and seed-source variation. Tree Physiol. 23, 685–694. doi:10.1139/x03-049

Mäkelä, A., Landsberg, J., Ek, A.R., Burk, T.E., Ter-Mikaelian, M., Ågren, G.I., Oliver, C.D., Puttonen, P., 2000. Process-based models for forest ecosystem management: current state of the art and challenges for practical implementation. Tree Physiol. 20, 289–298. doi:10.1093/treephys/20.5-6.289

Mathews, S., Bonser, S.P., 2005. Life histories, ecological tolerance limits, and the evolution of geographic range size in Eucalyptus (Myrtaceae). Aust. J. Bot. 53, 501–508. doi:10.1071/BT05010

140

McWilliams, W.H., Butler, B.J., Griffith, D.M., Hoppus, M.L., Lister, A.J., Lister, T.W., Morin, R.S., Stewart, L.B., Westfall, J.A., Steinman, J., Woodall, C.W., Caldwell, L.E., Lausten, K.M., Sader, S.A., Metzler, J.W., Williams, D.W., Whitman, A., 2005. The Forests of Maine : 2003, Resource Bulletin NE-164.

Medlyn, B.E., Duursma, R. a., Zeppel, M.J.B., 2011. Forest productivity under climate change: a checklist for evaluating model studies. Wiley Interdiscip. Rev. Clim. Chang. 2, 332–355. doi:10.1002/wcc.108

Meyer, S.R., 2009. Cooperative Forestry Research Unit 2009 Annual Report. University of Maine, Orono, Maine.

Mladenoff, D.J., 2004. LANDIS and forest landscape models. Ecol. Modell. 180, 7–19. doi:10.1016/j.ecolmodel.2004.03.016

Mohan, J.E., Cox, R.M., Iverson, L.R., 2009. Composition and carbon dynamics of forests in northeastern North America in a future, warmer worldThis article is one of a selection of papers from NE Forests 2100: A Synthesis of Climate Change Impacts on Forests of the Northeastern US and Eastern Canada. Can. J. For. Res. 39, 213–230. doi:10.1139/X08-185

Monserud, R.A., Leemans, R., 1992. Comparing global vegetation maps with the Kappa statistic. Ecol. Modell. 62, 275–293. doi:10.1016/0304-3800(92)90003-W

Morin, X., Lechowicz, M.J., 2013. Niche breadth and range area in North American trees. Ecography 36, 300–312. doi:10.1111/j.1600-0587.2012.07340.x

Moroni, M.T., Harris, D.D., 2011. Newfoundland balsam fir and black spruce orest described by the Newfoundland Forest Service Permanent Sample Plot and Temporary Sample Plots data sets. Information Report M-X-224El. Fredericton, NB.

Morris, A.E.C., Myerscough, P.J., 1991. Self-thinning intensity of nutrient over a gradient. J. Ecol. 79, 903–923.

Mosseler, A., Lynds, J., Major, J., 2003. Old-growth forests of the Acadian Forest Region. Environ. Rev. 77, 14–19. doi:10.1139/A03-015

Murphy, H.T., VanDerWal, J., Lovett-Doust, J., 2006. Distribution of abundance across the range in eastern North American trees. Glob. Ecol. Biogeogr. 15, 63–71. doi:10.1111/j.1466-822X.2006.00194.x

Newton, P.F., 2006. Asymptotic size-density relationships within self-thinning black spruce and jack pine stand-types: Parameter estimation and model reformulations. For. Ecol. Manage. 226, 49–59. doi:10.1016/j.foreco.2006.01.023

141

Newton, P.F., Smith, V.G., 1990. Reformulated self-thinning exponenets as applied to black spruce. Can. J. For. Res. 20, 887–893.

Nielsen, S.E., Johnson, C.J., Heard, D.C., Boyce, M.S., 2014. Can models of presence-absence be used to scale abundance? Two case studies considering extremes in life history. Ecography 28, 197–208.

Niinemets, Ü., Valladares, F., 2006. Tolerance to shade, drought, and waterlogging of temperate northern hemisphere trees and shrubs. Ecol. Monogr. 76, 521–547. doi:10.1890/0012-9615(2006)076[0521:TTSDAW]2.0.CO;2

Nitschke, C.R., Innes, J.L., 2008. A tree and climate assessment tool for modelling ecosystem response to climate change. Ecol. Modell. 210, 263–277. doi:10.1016/j.ecolmodel.2007.07.026

North East State Foresters Association, 2007. The Economic Importance and Wood Flows from Maine’s Forests, 2007.

Noss, R.F., 2001. Beyond Kyoto: Forest management in a time of rapid climate change. Conserv. Biol. 15, 578–590. doi:10.1046/j.1523-1739.2001.015003578.x

Noss, R.F., LaRoe, E, T., Scott, J.M., 1994. Endangered Ecosystems of the United States: A Preliminary Assessment of Loss and Degradation. Washington DC.

Nowacki, G.J., Abrams, M.D., 2014. Is climate an important driver of post-European vegetation change in the Eastern United States? Glob. Chang. Biol. 1–21. doi:10.1111/gcb.12663

Nunery, J.S., Keeton, W.S., 2010. Forest carbon storage in the northeastern United States: Net effects of harvesting frequency, post-harvest retention, and wood products. For. Ecol. Manage. 259, 1363–1375. doi:10.1016/j.foreco.2009.12.029

O’Hara, K.L., Ramage, B.S., 2013. Silviculture in an uncertain world: Utilizing multi-aged management systems to integrate disturbance. Forestry 86, 401–410. doi:10.1093/forestry/cpt012

Ollinger, S. V., Goodale, C.L., Hayhoe, K., Jenkins, J.P., 2007. Potential effects of climate change and rising CO2 on ecosystem processes in northeastern U.S. forests. Mitig. Adapt. Strateg. Glob. Chang. 13, 467–485. doi:10.1007/s11027-007-9128-z

Olson, D., Dinerstein, E., Wikramanyake, E., Burgess, N., Powell, G., Underwood, E., D’Amico, J., Itoua, I., Strand, H., Morrison, J., Loucks, C., Allnutt, T., Ricketts, T., Kura, Y., Lamoreux, J., Wettengel, W., Hedao, P., Kassem, K., 2001. Terrestrial ecoregions of the world: A new map of life on Earth. Bioscience 51, 933–938.

142

Oosting, H., Billings, W., 1951. A comparison of virgin spruce-fir forest in the northern and southern Appalachian system. Ecology 32, 84–103.

Park Williams, A., Allen, C.D., Macalady, A.K., Griffin, D., Woodhouse, C.A., Meko, D.M., Swetnam, T.W., Rauscher, S.A., Seager, R., Grissino-Mayer, H.D., Dean, J.S., Cook, E.R., Gangodagamage, C., Cai, M., McDowell, N.G., 2012. Temperature as a potent driver of regional forest drought stress and tree mortality. Nat. Clim. Chang. 3, 292–297. doi:10.1038/nclimate1693

Parmesan, C., 2006. Ecological and evolutionary responses to recent climate change. Annu. Rev. Ecol. Evol. Syst. 37, 637–669. doi:10.1146/annurev.ecolsys.37.091305.110100

Parmesan, C., Yohe, G., 2003. A globally coherent fingerprint of climate change impacts across natural systems. Nature 421, 37–42. doi:10.1038/nature01286

Pearson, R.G., Dawson, T.P., 2003. Predicting the impacts of climate change on the distribution of species: are bioclimate envelope models useful? Glob. Ecol. Biogeogr. 12, 361–371. doi:10.1046/j.1466-822X.2003.00042.x

Pearson, R.G., Dawson, T.P., Liu, C., 2004. Modelling species distributions in Britain : a hierarchical integration of climate and land-cover data. Ecography 3, 285–298.

Perala, D.A., Leary, R.A., Cieszewski, C.J., 1999. Self-thinning and stockability of the circumboreal aspens (Populus tremuloides Michx., and P. tremula L.) Res. Pap. NC-335. St. Paul, MN.

Pither, J., 2003. Climate tolerance and interspecific variation in geographic range size. Proc. Biol. Sci. 270, 475–481. doi:10.1098/rspb.2002.2275

Pontius, R., 2000. Quantification error versus location error in comparison of categorical maps. Photogramm. Eng. Remote Sensing 66, 1011–1016.

Porter, K.B., Maclean, D.A., Beaton, K.P., Upshall, J., 2001. New Brunswick permanent sample plot database (PSPDB v1.0): User’s guide and analysis, Information Report M-X- 209E Landscape Management Network. Canadian Forest Service - Atlantic Forestry Centre, Fredericton, NB.

Pöyry, J., Luoto, M., Heikkinen, R.K., Saarinen, K., 2008. Species traits are associated with the quality of bioclimatic models. Glob. Ecol. Biogeogr. 17, 403–414. doi:10.1111/j.1466-8238.2007.00373.x

Prasad, A.M., Iverson, L.R., Liaw, A., 2006. Newer classification and regression tree techniques: Bagging and random forests for ecological prediction. Ecosystems 9, 181– 199. doi:10.1007/s10021-005-0054-1

143

Pretzch, H., Biber, P., 2005. A re-evaluation of Reineke’s rule and stand density index. For. Sci.

Prince Edward Island Department of Agriculture and Forestry, 2002. State of the forest report 1990-2000. Planning for the future. Charlottetown, PI.

Puettmann, K.J., Hann, D.W., Hibbs, D.E., 1993. Evaluation of the size-density relationships for pure red alder and Douglas-fir stands. For. Sci. 39, 7–27.

Puettmann, K.J., Hibbs, D.E., Hann, D.W., 1992. The dynamics of mixed stands of Alnus rubra and Pseudotsuga menziesii: extension of size-density analysis to species mixture. J. Ecol. 80, 449–458.

R Core Team, 2013. R: A Language and Environment for Statistical Computing.

Rehfeldt, G., 2006. A spline model of climate for the western United States. USDA Forest Service, Rocky Mountain Research Statation, General Technical Report RMRD-GTR-165.

Reineke, L.H., 1933. Perfecting a stand-density index for even-aged forests. J. Agric. Res. 46, 627–638.

Ribbons, R.R., 2014. Disturbance and climatic effects on red spruce community dynamics at its southern continuous range margin. PeerJ 2:e293, 1–24. doi:10.7717/peerj.293

Rivoire, M., Le Moguedec, G., 2012. A generalized self-thinning relationship for multi- species and mixed-size forests. Ann. For. Sci. 69, 207–219. doi:10.1007/s13595-011- 0158-z

Rosenzweig, C., Karoly, D., Vicarelli, M., Neofotis, P., Wu, Q., Casassa, G., Menzel, A., Root, T.L., Estrella, N., Seguin, B., Tryjanowski, P., Liu, C., Rawlins, S., Imeson, A., 2008. Attributing physical and biological impacts to anthropogenic climate change. Nature 453, 353–357. doi:10.1038/nature06937

Royle, J., Nichols, J., 2003. Estimating abundance from repeated presence-absence data or point counts. Ecology 84, 777–790.

Sader, S.A., Bertrand, M., Wilson, E.H., 2003. Satellite change detection of forest harvest patterns on an industrial forest landscape. For. Sci. 49, 341–353.

Saunders, M.R., Seymour, R.S., Wagner, R.G., 2008. The Acadian Forest ecosystem research program: An example of natural disturbance-based silviculture in the Northeast, in: 60 Years of Research and Demonstration in Maine, 1950-2010. Gen. Tech. Rep. NRS-P- 123. U.S. Department of Agriculture, Forest Service, Northern Research Station, Newton Square, PA, pp. 71–80.

144

Scarl, J., 2012. Mountain Birdwatch 2.0: 2013. Annual Report to the United States Fish and Wildlife Service. Unpublished Report. Vermont Center for Ecostudies, Norwich, VT.

Schauffler, M., Jacobson, G.L., 2002. Persistence of coastal spruce refugia during the Holocene in northern New England, USA, detected by stand-scale pollen stratigraphies. J. Ecol. 90, 235–250.

Schwarz, P.A., Fahey, T.J., McCulloch, C.E., 2003. Factors controlling spatial variation of tree species abundance in a forested landscape. Ecology 84, 1862–1878. doi:10.1890/0012- 9658(2003)084[1862:FCSVOT]2.0.CO;2

Sea, W.B., Hanan, N.P., 2012. Self-thinning and tree competition in savannas. Biotropica 44, 189–196. doi:10.1111/j.1744-7429.2011.00789.x

Segurado, P., Araujo, M., 2004. An evaluation of methods for modelling species distributions. J. Biogeogr. 31, 1555–1568.

Seidl, R., Fernandes, P.M., Fonseca, T.F., Gillet, F., Jönsson, A.M., Merganičová, K., Netherer, S., Arpaci, A., Bontemps, J.-D., Bugmann, H., González-Olabarria, J.R., Lasch, P., Meredieu, C., Moreira, F., Schelhaas, M.-J., Mohren, F., 2011. Modelling natural disturbances in forest ecosystems: a review. Ecol. Modell. 222, 903–924. doi:10.1016/j.ecolmodel.2010.09.040

Seymour, R.S., 1992. The red spruce-balsam fir forest of Maine: Evolution of silvicultural practice in reponse to stand development patterns and disturbances, in: Ed., M.J.K. (Ed.), The Ecology and Silviculture of Mixed-Species Forests. pp. 217–244.

Seymour, R.S., 1995. The Northeastern Region, in: Barrett, J.W. (Ed.), Regional Silviculture of the United States. Wiley, New York, pp. 31–79.

Shaw, J.D., 2000. Application of stand density index to irregularly structured stands. West. J. Appl. For. 15, 40–42.

Shuman, B.N., Newby, P., Donnelly, J.P., 2009. Abrupt climate change as an important agent of ecological change in the Northeast U.S. throughout the past 15,000 years. Quat. Sci. Rev. 28, 1693–1709. doi:10.1016/j.quascirev.2009.04.005

Solomon, D.S., Zhang, L., 2002. Maximum size–density relationships for mixed softwoods in the northeastern USA. For. Ecol. Manage. 155, 163–170. doi:10.1016/S0378- 1127(01)00556-4

Sprugel, D.G., Bormann, F.H., 1981. Natural disturbance and the steady state in high-altitude balsam fir forests. Science 211, 390–393. doi:10.1126/science.211.4480.390

145

Stage, A., 1968. A tree-by-tree measure of site utilization for grand fir related to stand density index. U.S. Forest Service Research Note INT-77. Ogden, UT.

Stahl, U., Reu, B., Wirth, C., 2014. Predicting species ’ range limits from functional traits for the tree flora of North America 111, 13739–13744. doi:10.1073/pnas.1300673111

Stephenson, S., Adams, H., 1984. The spruce-fir forest on the summit of Mount Rogers in southwestern Virginia. Bull. Torrey Bot. Club 111, 69–75.

Sterba, H., Monserud, R.A., 1970. The maximum density concept applied to uneven-aged stands. For. Sci. 39, 432–453.

Stocker, T.F., Qin, D., Plattner, G., Alexander, L.V., Allen, S.K., Bindoff, N.L., Breon, F.-M., Church, J.A., Cubasch, U., Emori, S., Forster, P., Friedlingstein, P., Gillett, N., Greogry, J.M., Hartmann, D.L., Jansen, E., Kirtman, B., Knutti, R., Krishna Kumar, K., Lemke, P., Marotzke, J., Masson-Delmotte, V., Meehl, G.A., Mokhov, I.I., Piao, S., Ramaswamy, V., Randall, D., Rhein, M., Rojas, M., Sabine, C., Shindell, D., Talley, L.D., Vaughan, D.G., Xie, S.-P., 2013. 2013: Technical Summary, in: Stocker, T.F., Qin, D., Plattner, G., Tignor, M., Allen, S., Boschung, J., Nauels, A., Xia, Y., Bex, V., Midglet, P.M. (Eds.), Climate Change 2013: The Physical Science Basis. Contribution of Working Group I to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change. Cambridge University Press, Cambridge, United Kingdom and New York, NY, USA.

Stout, S.L., Nyland, R.D., 1986. Role of species composition in relative density measurement in Allegheny hardwoods. Can. J. For. Res. 16, 574–579. doi:10.1139/x86-099

Strobl, C., Boulesteix, A.-L., Zeileis, A., Hothorn, T., 2007. Bias in random forest variable importance measures: Illustrations, sources and a solution. BMC Bioinformatics 8:25. doi:10.1186/1471-2105-8-25

Strobl, C., Malley, J., Tutz, G., 2009. An introduction to recursive partitioning: rationale, application, and characteristics of classification and regression trees, bagging, and random forests. Psychol. Methods 14, 323–48. doi:10.1037/a0016973

Sturtevant, B.R., Bissonette, J.A., Long, J.N., 1998. Stand density management diagram for mixed balsam fir- black spruce stands. Database 15, 17–22.

Subedi, N., Sharma, M., 2013. Climate-diameter growth relationships of black spruce and jack pine trees in boreal Ontario, Canada. Glob. Chang. Biol. 19, 505–516. doi:10.1111/gcb.12033

Sutton, R.F., 1969. Silvics of white spruce (Picea glauca (Moench) Voss). Canada Forestry Service, Publication 1250, Ottawa, ON.

146

Swetnam, T.W., Betancourt, J.L., 1997. Mesoscale ecological responses to climatic variability in the American Southwest. J. Clim. 11, 3128–1347. doi:10.1007/978-90-481-8736- 2_32

Swift, E., Penner, M., Gagnon, R., Knox, J., 2007. A stand density management diagram for spruce in New Brunswick. For. Chron. 83, 187–197.

Tallent-Halsell, N.G. (Ed.), 1994. Forest Health Monitoring 1994 Field Methods Guide. EPA/620/R-94/097. Washington, D.C.

Tang, G., Beckage, B., 2010. Projecting the distribution of forests in New England in response to climate change. Divers. Distrib. 16, 144–158. doi:10.1111/j.1472- 4642.2009.00628.x

Taylor, A.R., Chen, Y.H., VanDamme, L., 2009. A review of forest succession models and their suitability for forest management planning. For. Sci. 1, 23–36.

Taylor, A.R., Wang, J.R., Kurz, W. a., 2008. Effects of harvesting intensity on carbon stocks in eastern Canadian red spruce (Picea rubens) forests: An exploratory analysis using the CBM-CFS3 simulation model. For. Ecol. Manage. 255, 3632–3641. doi:10.1016/j.foreco.2008.02.052

Thompson, I., Mackey, B., McNulty, S., Mosseler, A., 2009. Forest resilience, biodiversity, and climate change. A synthesis of the biodiversity/resilience/stability relationship in forest ecosystems, Secretariat of the Convention on Biological Diversity, Montreal. doi:10.1007/978-3-642-10445-9_5

Thompson, R.S., Anderson, K.H., Strickland, L.E., Shafer, S.L., Pelltier, R.T., Bartleain, P.J., 2006. Atlas of Relations between Climatic Parameters and Distributions of Important Trees and Shrubs in North America. U.S. Geological Survey Professional Paper 1650-D. doi:http://pubs.usgs.gov/pp/p1650-a/pages/conifers.html

Thomson, A.M., Riddell, C.L., Parker, W.H., 2009. Boreal forest provenance tests used to predict optimal growth and response to climate change: 2. Black spruce. Can. J. For. Res. 39, 143–153. doi:10.1139/X08-167

Tierney, G., Mitchell, B., Miller, K., Comiskey, J., Kozlowski, A., Faber-Langendoen, D., 2013. Northeast Temperate Network Long-term Forest Monitoring Protocol: 2012 revision. Natural Resource Report NPS/NETN/NRR 2013/639. National Park Service, Fort Collins, Colorado.

Tinner, W., Colombaroli, D., Heiri, O., 2013. The past ecology of Abies alba provides new perspectives on future responses of silver fir forests to global warming. Ecol. Monogr. 83, 419–439.

147

Townsend, P., 2004. Nova Scotia Forest Inventory Based on Permanent Sample Plots Measured between 1999 and 2003. Nova Scotia Natural Resources Forestry Division, Truro, N.S.

VanderSchaaf, C.L., Burkhart, H.E., 2007. Comparison of methods to estimate Reineke’s maximum size-density relationship species boundary line slope. For. Sci. 53, 435–442.

Vincent, A.B., 1965. Black spruce: A review of its silvics, ecology, and silviculture. Canada Department of Forests Forest Research Branch, Ottawa, ON.

Visser, H., de Nijs, T., 2006. The Map Comparison Kit. J. Environ. Model. Softw. 21, 346–358.

Volvfovicz Leon, R., 2011. Modelling the limiting size-density relationship of loblolly pine (Pinus taeda L.) Plantations.

Vuong, Q.H., 1989. Likelihood ratio tes for model selection and non-nested hypotheses. Econometrica 57, 307–333.

Wake, C., Burakowski, L., Lines, G., McKenzie, K., Huntington, T., 2006. Cross border indicators of climate change over the past century: Northeastern United States and Canadian Maritime Region, The Climate Change Task Force of the Gulf of Maine Council on the Marine Environment in cooperation with Environment Canada and Clean Air-Cool Planet. Buscow, NH.

Walter, L.C., 1967. Silviculture of the minor southern conifers. Stephen F. Austin State College School of Forestry, Bulletin 15, Nacogdoches, TX.

Weiskittel, A., Gould, P., Temesgen, H., 2009. Sources of variation in the self-thinning boundary line for three species with varying levels of shade tolerance. For. Sci.

Weiskittel, A.R., Hann, D.W., Kershaw, J.A., Vanclay, J.K., 2011. Forest growth and yield modeling. Wiley-Blackwell, West Sussex, UK.

Weller, D.E., 1990. Will the real self-thinning rule please stand up?--A reply to Osawa and Sugita. Ecology 71, 1204–1207.

Westveld, M., 1931. Reproduction on pulpwood lands in the Northeast. USDA Tech. Bull. 223. Washington, D.C.

Westveld, M., 1946. Forest management as a means of controlling the spruce budworm. J. For. 44, 949–953.

Williams, J.W., Kharouba, H.M., Veloz, S., Vellend, M., Mclachlan, J., Liu, Z., Otto-Bliesner, B., He, F., 2013. The ice age ecologist: Testing methods for reserve prioritization during

148

the last global warming. Glob. Ecol. Biogeogr. 22, 289–301. doi:10.1111/j.1466- 8238.2012.00760.x

Williams, R.A., 2003. Use of stand density index as an alternative to stocking percent in upland hardwoods. North. J. Appl. For. 20, 137–142.

Wilson, D.S., Seymour, R.S., Maguire, D. a, 1999. Density management diagram for Northeastern red spruce and balsam fir forests. North. J. Appl. For. 16, 48–56.

Woodall, C.W., Miles, P.D., Vissage, J.S., 2005. Determining maximum stand density index in mixed species stands for strategic-scale stocking assessments. For. Ecol. Manage. 216, 367–377. doi:10.1016/j.foreco.2005.05.050

Yang, Y., Titus, S.J., 2002. Maximum size-density relationship for constraining individual tree mortality functions. For. Ecol. Manage. 168, 259–273. doi:10.1016/S0378- 1127(01)00741-1

Yoda, K., Kira, T., Ogawa, H., Self-thinning, K.H., 1963. Self-thinning in overcrowded pure stands under cultivated and natural conditions. J. Biol. Osaka City Univ. 14, 1963.

Yu, K., Zhang, J., 2005. A three-parameter asymmetric Laplace distribution and its extension. Commun. Stat. - Theory Methods 34, 1867–1879. doi:10.1080/03610920500199018

Zeide, B., 1983. The mean diameter for stand density index. Can. J. For. Res. 13, 1023–1024. doi:10.1139/x83-135

Zeide, B., 1987. Analysis of the 3/2 power law of self-thinning. For. Sci. 33, 517–537.

Zeide, B., 2005. How to measure stand density. Trees 19, 1–14. doi:10.1007/s00468-004- 0343-x

Zeileis, A., Kleiber, C., Jackman, S., 2007. Regression models for count data in R. J. Stat. Softw. 27, 1–25.

Zhang, J., Oliver, W.W., Powers, R.F., 2013. Reevaluating the self-thinning boundary line for ponderosa pine (Pinus ponderosa) forests. Can. J. For. Res. 971, 963–971. doi:10.1139/cjfr-2013-0133

Zhang, L., Bi, H., Gove, J.H., Heath, L.S., 2005. A comparison of alternative methods for estimating the self-thinning boundary line. Can. J. For. Res. 35, 1507–1514. doi:10.1139/x05-070

Zielinski, G., Keim, B., 2003. New England Weather, New England Climate. University Press of New England.

149

APPENDICES

150

APPENDIX A: Data Sources

Québec Permanent Sample Plots

Tree data was obtained from a variety of Permanent Sample Plots (PSPs) in Québec.

The majority of the data was obtained from the Québec Ministère des Ressources naturelles. Data was collected between 1970 and 2011 from 0.04 ha plots that were remeasured on average at five year intervals. Other data sources included the Fédération des producteurs de bois du Québec, Parks Canada, Service de la Comptablité Forestière,

Service de la protection des insectes et des maladies, and University of Laval. Data collection began at different times between 1970 and 1996 and continued until 2008. Mean plot sizes across these plots ranged from 0.32 to 0.40 ha in size and mean measurement intervals ranged from five to 11 years (Li et al., 2011).

Nova Scotia Permanent Sample Plots

Individual dbh measurements were obtained from 3,230 PSPs from Nova Scotia’s

Department of Natural Resources Forestry Division. All PSPs were 0.04 ha in size (Townsend,

2004). Tree dbh began measurement in 1965, and measurement intervals averaged five years.

New Brunswick Permanent Sample Plots

Data was collected from PSPs managed by the New Brunswick Department of

Natural Resources. Plot sizes varied by density (Porter et al., 2001). The majority of the data came from 1,769 0.04 ha plots, while the remaining 688 plots varied from 0.0008 to 1 ha in

151 size. Data collection began in 1985 and was remeasured on approximately five year intervals.

Newfoundland Permanent Sample Plots

PSP data was collected from 1,003 plots maintained by the Newfoundland Forest

Service. Plot size varied by density and ranged from 0.002 to 0.1 ha, but the majority of data was collected from 0.04 ha plots. Plot sampling was initiated in 1985 and remeasured on four to five year intervals (Moroni and Harris, 2011).

Penobscot Experimental Forest

Long-term tree data was obtained from numerous studies that occurred at the USFS

Penobscot Experiment Forest (PEF), located in the towns of Bradley and Eddington, Maine.

Continuous forest inventory overstory tree data was collected from 248 0.02 ha plots beginning in 1974 and recollected on an average of five year intervals (Russell et al., 2014).

Additional data came from 295 0.008 ha plots collected between 1976 and 2008 as part of a long term pre-commercial thinning (PCT) study and 180 0.01 ha plots from the Acadian

Forest Ecosystem Research Program (AFERP), remeasured on average of three times between 1995 and 2008 (Saunders et al., 2008).

Cooperative Forestry Research Unit

Individual tree measurements were acquired from three sources throughout Maine managed by the University of Maine’s Cooperative Forestry Research Unit (CFRU). Data was collected once from Austin Pond in Somerset, Maine in 1999 on 26 0.021-ha plots as part of a long term study examining the effects PCT and herbicide treatment on spruce-fir

152 regeneration. Additional data was collected from a thinning study on 31 1 ha plots between

1978 and 1994 in Northern Maine. Lastly, data was acquired from the commercial thinning research network (CTRN). The CTRN data includes twelve research location across Maine monitoring the effects PCT in spruce and fir stands. Tree measurements began in 2001 and were remeasured annually or biannually through 2010 on 0.08 ha plots (Meyer, 2009).

University of Maine Research

Research completed by associates of the University of Maine was supplied to assist in analysis. Dr. Sean Fraver shared data from 34 0.15-ha and three 0.25-ha plots collected at

Big Reed Forest Reserve located in northern Piscataquis County, Maine (Fraver et al., 2007) and Dr. Thomas Brann supplied data from 424 0.02 ha plots revisited on an annual basis between 1974 to 1985.

Prince Edward Island Permanent Inventory Plots

The Prince Edward Island Department of Agriculture and Forestry maintains a network of 803 forested Permanent Inventory Plots (PIP), established in 1999 (Prince

Edward Island Department of Agriculture and Forestry, 2002). Plot size was unknown at the time of analysis, and data was only used in presence/absence analyses.

New Hampshire Division of Forests and Lands

Data was collected from New Hampshire’s Forest Health Monitoring (FHM),

Continuous Forest Inventory (CFI), and Growth Point programs. FHM protocols are established nationally (Tallent-Halsell, 1994). Data was collected annually from 2003 – 2013 at three different locations throughout the state. At each location, tree data was collected from four 0.016 ha plots located within 36.6 meters from one another. CFI data was

153 obtained for the Caroline A. Fox Research and Demonstration Forest. Approximately 68

16.03 m radius plots have been monitored since 1955 at approximately 10-year intervals.

Lastly, five growth points were established in Honey Brook State Forest in 2013 to track growth in a red spruce habitat. Sample trees were determined using a 20 basal area factor

(BAF) prism.

Vermont Monitoring Cooperative

FHM Data was collected from the Lye Brook Wilderness Area (Green Mountain

National Forest) and Mt. Mansfield State Forest in Vermont. Data is managed by the

Vermont Monitoring Cooperative, a partnership by the State of Vermont, the University of

Vermont and the USDA Forest Service, that manages forest ecosystem data. Tree data was collected from four 0.016 ha plots at 20 different locations throughout the two areas at approximately annual intervals.

National Park Service - Northeast Temperate Network

The Northeast Temperate Network consists of eleven parks owned by the National

Park Service in the northeastern United States. The largest park in this network, Acadia

National Park, is primarily spruce-fir habitat (Tierney et al., 2013). Since 2006 individual tree measurements have been collected at four year intervals on 176 0.0225 ha plots. Additional data was collected from 174 0.04 ha plots spread throughout seven smaller national historic parks and national historic sites in the network. These additional plots primarily consist of northern hardwood and central hardwood habitat.

154

Vermont Center for Ecostudies

Tree data was collected from the Vermont Center for Ecostudies’ Mountain

Birdwatch program to target high elevation spruce-fir habitat. These datasets include plots located in the Adirondack Mountains of New York, the Green Mountains of Vermont, the

White Mountains of New Hampshire and Maine, and the Appalachian Mountains in northern Maine. Mountain Bird Watch established 131 transects between 2010 and 2011.

These sites were re-measured on an annual basis. Each transect consisted of between 3 and

6 plots located 250 m from one another. A ten BAF wedge prism was used to count tree by species present at each plot (Scarl, 2012).

University of Massachusetts

Data from the research of Dr. William DeLuca at the University of Massachusetts-

Amherst was collected to target high elevation spruce-fir populations. This data was collected following two different protocols. Individual tree dbhs were collected from 42 0.04 ha plots in Vermont and New York in 2011 and 2012. In New Hampshire, individual species composition was measured as a percent of total canopy make up at 127 plots. The data from New Hampshire was used for presence/absence analysis only (Deluca and King, 2014).

Massachusetts Department of Conservation and Recreation

Individual tree dbh measurements were obtained from the Massachusetts

Department of Conservation and Recreation’s continuous forest inventory for the Quabbin

Reservoir watershed located in Massachusetts. In 1960, 347 0.08 ha plots were established and remeasured on a five or ten-year basis (Kyker-Snowman et al., 2007). Five plots with spruce-fir habitat were made available for these analyses

155

Table A.1. Description of different data sources used in analyses. % of Plots Remeasurement Geographic Number of Measurement Plot Size with Source Owner Number of Interval Observations a Region Plots (years) Period (ha)/Prism Spruce or Fir Forest Inventory and US Forest Service Eastern US 6,833,159 194,838 Varies 1968-2010 0.07b 0.50% Analysis (FIA) Southern Québec PSP Québec Ministry of 1,583,176 39,436 5 1970-2013 0.04 84.5 Natural Resources Québec Nova Scotia Department of Nova Scotia PSP Nova Scotia 494,108 3,042 5 1965-2006 0.04 94.7 Natural Resources Forestry Division New Brunswick New

New Brunswick PSP 493,104 2,387 5 1985-2005 94.1

Page Department of Brunswick Natural Resources. 0.04c

156 Southeast Québec Research PSP Québec Ministry of 3,069 5 to 11 1970-2008 0.32 - 0.40 88.7 Québec Natural Resources 321,855 Newfoundland Forest Newfoundland PSP Newfoundland 321,550 1,291 4 or 5 1985-2008 0.04d 100 Service Penobscot US Forest Service Central Maine 169,118 562 Varies 1974-2008 Varies 98.2 Experimental Foreste Commercial Thinning Cooperative Forestry Northern ME 78 1 or 2 2000-2007 0.08 100 Research Network Research Unit 80,035 Northern Brann GIS University of Maine 365 1 1975-1985 0.04 100 Maine 64,570 0.01 or AFERP University of Maine Central Maine 180 5 1995-2007 98.9 31,850 0.05 Prince Edward Island Prince Edward Island Department of Prince Edward 26,782 691 - 1999 - ? - 91.3 PSP Agriculture and Island Forestry

Table A.1. continued % of Plots Remeasurement Geographic Number of Measurement Plot Size with Source Owner Number of Interval Observations a Region Plots (years) Period (ha)/Prism Spruce or Fir New Hampshire Caroline A. Fox Southern New Division of Parks and 20,118 65 10 1955-2011 0.08 33.3 Research Forest Hampshire Lands Vermont Forest Vermont Monitoring Vermont 17,065 76 1 1992-2013 0.06 63.2 Health Monitoring Cooperative Northeast Temperate Northeastern 0.02 or National Park Service 14,532 324 4 2006-2013 40.7 Network US 0.04f Cooperative Forestry Central Maine 10,267 207 - 1999 0.02 100 Austin Pond Research Unit High

Page Mountain Birdwatch Vermont Center for elevations in 10 BAF 5,797 2,008 1 2010-2011 99.4 Program Ecostudies New England prism 157 and New York

Big Reed Forest 0.15 or University of Maine Central Maine 3,102 37 - 2000-2001 97.3 Reserve 0.25

New Hampshire New Hampshire New Forest Health Division of Parks and 2,939 16 1 2003-2013 0.06 100 Hampshire Monitoring Lands

High High Elevation Bird University of elevations in 1,752 151 1 2011-2013 0.04 94.7 Habitat Massachusetts New England and New York Database maintained New England Witness Tree Data 1,342 778 - 1623-1859 NA 72.6 by Charles Cogbill and New York

Table A.1.continued % of Plots Remeasurement Geographic Number of Measurement Plot Size with Source Owner Number of Interval Observations a Region Plots (years) Period (ha)/Prism Spruce or Fir McCormack Thinning Cooperative Forestry Northern 691 14 NA 1978-1994 1 100 Study Research Unit Maine Massachusetts Department of Central Quabbin Reservoir CFI 456 5 5 or 10 1960-2010 0.08 80 Conservation and Massachusetts Recreation

New Hampshire Southern 20 BAF HoneyBrook Division of Parks and NNew 38 5 - 2013 100 prism Lands Hampshire

Page

158 a Majority or most frequent plot sizes reported

b Sampling design for FIA implemented in 1998. Prior to this data sampling designs varied by region and were taken into account in analyses.

c Plot size varied by tree density. 80% of plots were 0.04 ha in size. The remaining 29% varied from 0.0008 to 0.02 ha in size (NB)

d Plot size varied by tree density. 34% of plots were 0.04 ha in size. The remaining 66% varied from 0.1 to 1 ha in size

e Data from numerous studies within the Penobscot Experimental Forest were used including a continuous forest inventory (CFI), a long term pre-commercial thinning study (PCT), and the research of Dr. Mike Saunders.

f 0.02 ha plots at Acadia National Park. 0.04 at all other National Parks in the Network.

APPENDIX B: Effect of Tree Diameter Thresholds on Analysis

Table B.1. Results of random forest analyses for presence/absence modeling performed with a threshold of 1 cm and 5 cm as a requirement for individuals included in analysis. The prevalence ratio is a ratio of prevalence to an absence sample from within the hypervolume (HV) to an absence sample from outside the HV. OOB = Out of bag; AUC = Area under receiver operator curve. Prevalence OOB Species Specificity Sensitivity AUC Top 5 Variables Ratio Error

THRESHOLD OF 1 CM

PRDD5, PRMTCM, MAPMTCM, Balsam Fir 55-20-25 5.18 92.26 96.92 0.98 MAPDD5, GSPMTCM

White PRDD5, PRMTCM, MAPMTCM, 50-25-25 3.89 92.71 99.51 0.98 Spruce MAPDD5, GSPMTCM

Black MAPDD5, PRMTCM, PRDD5, 55-20-25 4.37 91.91 99.54 0.99 Spruce MAPMTCM, GSPMTCM

PRMTCM, PRDD5, MAPDD5, Red Spruce 40-40-20 3.00 95.21 99.67 0.99 MAPMTCM, GSPMTCM

THRESHOLD OF 5 CM

PRDD5, MAPDD5, PRMTCM, Balsam Fir 55-20-25 3.31 94.07 98.82 0.98 GSPMTCM, MAPMTCM

White PRDD5, PRMTCM, MAPMTCM, 50-25-25 4.03 92.42 99.52 0.98 Spruce MAPDD5, GSPMTCM

Black PRMTCM, MAPDD5, PRDD5, 55-20-25 4.17 91.33 99.52 0.99 Spruce MAPMTCM, GSPMTCM

PRMTCM, PRDD5, MAPMTCM, Red Spruce 40-40-20 3.16 95.15 99.38 0.99 MAPDD5, GSPMTCM

159

Figure B.1. Mapped predictions of presence/absence models using a data inclusion threshold of 1 cm and 5 cm for balsam fir.

160

Figure B.2. Mapped predictions of presence/absence models using a data inclusion threshold of 1 cm and 5 cm for white spruce.

161

Figure B.3. Mapped predictions of presence/absence models using a data inclusion threshold of 1 cm and 5 cm for black spruce.

162

Figure B.4. Mapped predictions of presence/absence models using a data inclusion threshold of 1 cm and 5 cm for red spruce.

163

APPENDIX C: Effect of Solely Using Forest Inventory and Analysis Data for Acadian Forest

Spruce-Fir Species Distribution Models

Presence/absence models were generated for balsam (Abies balsamea L.), white spruce (Picea glauca (Moench) Voss), black spruce (Picea mariana (Miller) B.S.P.), and red spruce (Picea rubens Sarg.) using only Forest Inventory and Analysis (FIA) data from the

United States (U.S.) Forest Service (USFS). Results are presented in Table C.1. The mapped prediction objects for each species are presented in Figure C.1. Overall, FIA models were able to predict species’ distributions well within the U.S., but were unable to accurately portray species’ ranges on unknown surfaces in Canada. Within the U.S., balsam fir was likely overpredicted in the Adirondacks, and white spruce on the Pennsylvania and New

York border. Black spruce was falsely predicted as vastly present in the Adirondacks and over represented in Maine. Balsam fir and white spruce habitats were grossly overpredicted in Canada, while much of black spruce’s habitat was missed. Red spruce’s range was falsely extended into parts of Québec and Newfoundland.

FIA data is an uniformly generated unbiased dataset that is considered representative of the landscape in the U.S. (Bechtold and Patterson, 2005). Using solely FIA data to model the species of interest in the study did not generate accurate results beyond the perimeter of the United States. FIA data does have potential in modeling species’ distribution that are bounded within the U.S. For example, studies performed at a broad resolution (Iverson et al., 2008) or studies of species that were contained within the U.S.

(Joyce and Rehfeldt, 2013) have had good results.

164

Depiction of black spruce using only FIA data was poor. This is likely due to a low number of examples of species presence. Taking into account knowledge of black spruce distribution within the U.S., supported by additional data collected for this study, it appears that FIA data collection was unable to capture species occurrence within Maine. Predictions generated with this data overpredicted current distribution in Maine, as well as in upstate

New York. The absence of data points given by the FIA data in general, led to overprediction as opposed to under representation. This is in part due to model construction, but is also representative of the fact that suffering from lack of adequate data to fully characterize species-climate interactions will results in the inability to realize species-niche limitations, rather than miss areas of habitat appropriateness. While FIA data has limitations, it should not necessarily be compared in quality to the additional data used in the study, as this data was largely selected for the presence of spruce and fir.

Table C.1. Results for presence/absence modeling with only US Forest Service Forest Inventory and Analysis data. OOB = Out of bag; AUC = Area under receiver operator curve. OOB Species Specificity Sensitivity AUC Top 5 Variables Error PRDD5, MAPTD, MAPMTCM, GSPMTCM, Balsam Fir 2.0 95.6 99.9 0.99 PRMTCM White MTCMGSP, MAPMTCM, MTCMMAP, 2.6 94.7 100.0 0.99 Spruce GSPMTCM, MAPDD5 Black PRDD5, MTCMGSP, MTCMMAP, TDGSP, 5.3 88.3 99.9 0.98 Spruce MAPMTCM MAPMTCM, GSPMTCM, MTCMGSP, MAPDD5, Red Spruce 3.3 95.1 99.0 0.99 MTCMMAP

165

Figure C.1. Mapped predictions objects for presence/absence models for each species generated with solely United States Forest Service Forest Inventory and Analysis data.

166

APPENDIX D: Testing the Output of Likelihood Models as a Predictor of Abundance

To determine if a higher likelihood of occurrence translates to more abundance, indicating the core of distribution, models were fit between the two random forest ouputs for balsam fir (Abies balsamea L.), white spruce (Picea glauca (Moench) Voss), black spruce

(Picea mariana (Miller) B.S.P.), and red spruce (Picea rubens Sarg.). Modeling abundance with presence/absence data has been shown possible, dependent on species’ relationship with the environment (Barry and Welsh, 2002; Nielsen et al., 2014; Royle and Nichols,

2003). The probability prediction objects of the presence/absence models were compared to predicted abundance. Only the relative basal area (BA) abundance metric was used for these analyses. Prediction objects for both likelihood and abundance estimates are of the same size, and thus every pixel in the prediction matrices were assigned both a likelihood value and an abundance value. This facilitated direct comparison with model fitting. The large proportion of absences in the predicted datasets necessitated the use of models that do not rely on the assumptions of normal distribution. Models considered in this analysis included a generalized linear model (GLM), a zero-inflated regression model (ZIM), and a zero-altered model (ZAM) each with a negative binomial distribution.

A negative binomial distributed accounts for over dispersion in the data set that arises from the implicit heterogeneity of tree composition across the large landscape used in this analysis. At this scale the majority of data is concentrated in absence or low numbers across the landscape, reflecting non-ideal habitat or the influence of competition and disturbance on species occurrence, with select spots of high species abundance. This results

167

in a low mean and a high variance that exceeds the mean. The negative binomial distribution accounts for this over dispersion with an additional parameter, theta (k).

Distribution of the model is defined as (Lawless, 1987; Li et al., 2011):

NB(y)= Γ(y+1k)Γ(1k)y!(1μk+1)k(μkμk+1)y

Where y is the random variable, µ is the mean, and Γ represents the Gamma distribution.

Variance is defined as Var(y)= μ+μ2k. When k exceeds 10 the distribution behaves like a

Poisson distribution. The negative binomial can be viewed as an overdispersed Poisson, where the k parameter of the Poisson is exhibiting a Gamma distribution (Royle and Nichols,

2003). ZIM and ZAM models improve upon the typical GLM in this scenario by dividing and fitting the data in two parts; one that accounts for the zeroes in the data and one that accounts for values above zero. The difference between ZIM and ZAM is subtle and lies in how the zeroes are modeled. In a ZIM model, zero data is divided into two parts: those caused by a binomial mechanism and those caused by negative binomial distribution. ZAM accounts for all zeroes through a binomial process (Zeileis et al., 2007). Models fits were compared via Akaike information criterion (AIC) and -2log-likelihood (-2logL) and assessed for accuracy by comparing them to actual distributions. Smaller values of AIC and -2logL indicate a better fit.

Negative binomial distribution modeling exhibited limited success in describing the relationship between the two prediction objects. Zeros composed on average of 56% of the observed frequency of the abundance model outputs. Average mean (± SD) ranged from

3.7% (± 8.3) for red spruce to 22.6% (± 30.2) for black spruce. The average observed variance to mean ratio for the response variable ranged from 12.5% (P. glauca) to 40.2%

168

(P.mariana) which suggests over dispersion in the data. The AIC and -2logL indicated that the GLM negative binomial performed substantially worse than those that incorporated a second regression for zeros into their model form (Table D.1). ZIM and ZAM performed similarly, with the AIC and -2logL demonstrating ZIM performed marginally better in most cases. The Vuong (1989) hypothesis test, designed for non-nested models, confirmed that

ZIM was the better fit for all models (p<0.0001). Coefficients are the ZIM models are shown in Table D.2.

Both ZIM and ZAM were able to capture similar zero frequencies when compared to the actual model outputs (Table D.3), indicating that most of the zeros were captured by modeling through a binomial process. The ZIM was able to precisely describe the mean of the observed datasets (1.4% average percent difference), but failed to capture the full variance. On average, the variance to mean ratio differed by 29.4%. The failure to capture the full effect of the variance exhibited itself by over representing values below or close to the mean and underestimating or completely missing values concentrated at the higher range of values.

It was difficult to capture high levels of abundance with negative binomial models.

Negative binomial regression is typically conserved for count data. While BA can be considered count data, the data was weighted as a proportion prior to abundance modeling.

It is possible this weighting concentrated values in an unnatural dispersion form, affecting model performance. Furthermore, model performance seems to be affected by low distribution of values in the upper range of the dataset. For example, the red spruce ZIM failed to capture values greater than 40%, but the observed values above this mark only

169

Table D.1. Values for negative binomial model comparison models. GLM = generalized linear model; ZAM = zero adjusted model; ZIM = Zero inflated model; -2logl = -2log-likelihood; AIC = Akaikie information criterion. Species Model Form -2logL AIC Balsam Fir GLM 10830930 21661865 ZAM 8756949 17513908 ZIM 8669185 17338380 White Spruce GLM 10871139 21742283 ZAM 8387052 16774113 ZIM 8240570 16481151 Black Spruce GLM 12133382 24266770 ZAM 9444979 18889967 ZIM 9390209 18780429 Red Spruce GLM 7115475 14230949 ZAM 5868185 11736381 ZIM 5735562 11471133

composed 0.8% of the dataset. Similarly, values missed for white spruce composed 2.7% of the dataset, 5.5% for balsam fir, and 12.8% for black spruce. While the percent of the values missed is low, capturing these values is important as they represent suitable habitat for the species of the spruce-fir forest. It is important to note that the abundance model output underpredicted high BA values and this error affected, and was further compounded, in the negative binomial models.

170

Table D.2. Coefficients for zero-inflated model (ZIM) for each species.SE = standard error. Species Model Parameter Estimate SE p-value

Balsam fir y0 3.7510 0.0007 <0.0001

β0 -0.0198 0.0000 <0.0001

log(k) 0.9532 0.0011 <0.0001

y1 -0.0166 0.5662 <0.0001

β1 1.7010 0.0057 <0.0001

White spruce y0 2.1830 0.0007 <0.0001

β0 0.0086 0.0000 <0.0001

log(k) 1.0980 0.0012 <0.0001

y1 2.6724 0.0029 <0.0001

β1 -2.7399 0.0078 <0.0001

Black Spruce y0 4.2140 0.0004 <0.0001

β0 -0.0220 0.0000 <0.0001

log(k) 1.4920 0.0012 <0.0001

y1 -0.0176 0.5932 <0.0001

β1 1.8010 0.0060 <0.0001

Red Spruce y0 3.4300 0.0013 <0.0001

β0 -0.0208 0.0000 <0.0001

log(k) 0.6414 0.0063 <0.0001

y1 -0.0202 0.3912 <0.0001

β1 2.0490 0.0039 <0.0001

171

Table D.3. Observed versus predicted frequencies for negative binomial models for each species. GLM = generalized linear model; ZAM = zero adjusted model; ZIM = Zero inflated model Observed Predicted Frequency Species Range Frequency NB ZIM ZAM

Balsam Fir 0 2138569 2222677 2097706 2134697

1-10 493231 572317 363913 303972

11-20 559485 215394 431809 454759

21-30 404297 163867 440856 440856

31-40 259899 160029 591783 656687

41-50 156548 163038 333668 268764

51-60 88479 128784 0 0

61-70 61850 142210 0 0

71-80 54181 157751 0 0

81-90 24802 215683 0 0

91-100 5031 117985 0 0

White 0 2186701 2553688 2092168 219002 Spruce 1-10 783741 706221 599708 362686

11-20 886251 307853 1426602 1618871

21-30 29402 218536 141257 87176

31-40 78240 215110 0 0

41-50 21091 57922 0 0

51-60 6813 57922 0 0

61-70 2247 0 0 0

>70 559 0 0 0

Black 0 2155749 2194834 2158821 2158821 Spruce 1-10 291822 462039 172749 162379

11-20 216468 130407 273687 273688

21-30 223542 90836 160112 170481

31-40 210574 81204 164184 153060

172

Table D.3. continued Observed Predicted Frequency Species Range Frequency NB ZIM ZAM

Black 41-50 195541 87706 178824 189948 Spruce 51-60 217038 79638 231239 231239

61-70 234832 82608 920119 920119

71-80 239726 55641 0 0

81-90 184099 130800 0 0

91-100 90344 864022 0 0

Red Spruce 0 2978780 3527191 2980604 3111969

1-10 719732 247957 761993 619037

11-20 323026 80812 203866 215457

21-30 137379 60267 252187 278396

31-40 63005 47043 61085 34876

41-50 26244 50330 0 0

51-60 8379 32496 0 0

61-70 2495 58356 0 0

71-80 636 34167 0 0

>80 56 121116 0 0

173

BIOGRAPHY OF THE AUTHOR

Caitlin Marie Bernadette Andrews was born in Thornwood, New York to Joseph and

Margaret Andrews, and is the youngest of four sisters. She graduated from Our Lady of the

Good Counsel Academy in White Plains, NY in 2004 before escaping to the wilds of New

England where she graduated from the University of Vermont with a B.S. in Environmental

Science in 2009. She explored the U.S. and parts of the world, working in California,

Washington, Indiana, and Thailand, before deciding to pursue her M.S. in Forest Resources at the University of Maine in 2012. She currently resides in Big Bend National Park in Texas with her fiancé and pet family. She is a candidate for the Master of Science degree in Forest

Resources from the University of Maine in December 2016.

174