EXAMINING THE DISTRIBUTION OF DICYMBE CORYMBOSA MONODOMINANT FORESTS IN WESTERN GUYANA USING SATELLITE IMAGERY
by Rebecca S. Degagne A Thesis Presented to The Faculty of Humboldt State University In Partial Fulfillment Of the Requirements for the Degree Master of Science In Natural Resources: Natural Resources Planning and Interpretation August, 2007
EXAMINING THE DISTRIBUTION OF DICYMBE CORYMBOSA MONODOMINANT FORESTS IN WESTERN GUYANA USING SATELLITE IMAGERY by Rebecca S. Degagne Approved by the Master’s Thesis Committee: Steven Steinberg, Major Professor Date Terry W. Henkel, Committee Member Date Lawrence Fox III , Committee Member Date Coordinator, Natural Resources Graduate Program Date 07 PI 632 07/30 Natural Resources Graduate Program Number Chris A. Hopper , Interim Dean for Research and Graduate Studies Date
ii
ABSTRACT Examining The Distribution Of Dicymbe corymbosa Monodominant Forests In Western Guyana Using Satellite Imagery Rebecca S. Degagne The ectomycorrhizal canopy tree Dicymbe corymbosa (Caesalpiniaceae) forms single species dominated ( i.e. monodominant) forests in the central Pakaraima Mountains of western Guyana. Dicymbe corymbosa , like other tropical monodominant species, has important life history traits that promote conspecific clumping, the study of which may allow a better understanding of the phenomenon of high tree species diversity characterizing much of the tropics, and how it is circumvented. Dicymbe corymbosa forests, occurring in Guyana as patches within a non ectomycorrhizal, mixed species forest matrix, function as habitat islands for a diverse assemblage of putatively endemic ectomycorrhizal fungi. Ground based studies have not adequately determined the regional extent of D. corymbosa forests, nor are they practical. Distribution information is critical to allow broader sampling of these forests and their ectomycorrhizal fungal constituents and ultimately to inform conservation plans for these unique habitats. The rugged, remote nature of the study site and spatially discrete occurrence of D. corymbosa stands suggest satellite image analysis as an ideal tool for determining the extent of these relatively unknown tropical forests. We assessed the suitability of Landsat satellite data for mapping regional distribution of D. corymbosa forests in Guyana’s Upper Potaro
River Basin. Supervised image classification (maximum likelihood algorithm) was performed on images from 9 August 1989 (Landsat 5 TM) and 16 October 1999
(Landsat 7 ETM+). In situ forest reference data were used to quantitatively access
iii accuracy of output classification maps. Supervised classification of Landsat 5 TM and
Landsat 7 ETM+ imagery performed well in distinguishing forest types dominated by a single species from mixed species forest. For both the 1989 and 1999 images, D. corymbosa forest class accuracy is acceptable (User’s accuracy = 89.8 %, Khat = 0.74
(1989); User’s accuracy = 80.7 %, Khat = 0.59 (1999). Results indicate my output classification maps will be useful in developing efficient sampling schemes for future forest ecology studies and field surveys for ectomycorrhizal fungi. Supervised classification of Landsat data may also be effective for identifying monodominant forests in other remote regions of the tropics.
iv
ACKNOWLEDGEMENTS
Financial support was provided by Humboldt State University Foundation’s
Research, Scholarship, and Creative Activities program, and the Federal Work Study graduate research program. Research permits were granted by the Guyana
Environmental Protection Agency. Field assistance was provided by Valentino Joseph,
Peter Joseph, Chris Andrew, Luciano Edmond, Lance Woolley, and Luke Henault.
Landsat images were made available by The Global Land Cover Facility (University of
Maryland, www.landcover.org). My advisors and other faculty at Humboldt State
University provided invaluable support and guidance, as did my friends and family.
v
TABLE OF CONTENTS
Page
ABSTRACT ………………………………………..…………………………………... iii
ACKNOWLEDGEMENTS ………………………………………………………...…… v
LIST OF TABLES ……………………………………………………………...…….... vii
LIST OF FIGURES ………………………………………………………………...…. viii
INTRODUCTION ………………………………………………………………….…… 9
STUDY SITE ……………………………………………………………...…………… 14
MATERIALS AND METHODS …………………………………………………….… 16
Satellite Data ………………………………………………………………...…. 16
Preliminary Data Processing ………………………………………………...…. 16
Classification Procedure …………………………………………………...... … 20
Thematic Accuracy Assessment ……………………………………………..… 21
RESULTS ………………………………………………………………………..…….. 22
DISCUSSION ……………………………………………………………………..…… 26
Classification of Forest Types …………………………………………………. 26
Regional Distribution ………………………………………………………...… 28
Future Applications ………………………………………………………..…… 30
LITERATURE CITED ……………………………………………………………….... 32
vi
LIST OF TABLES
Table Page
1 Characteristics of vegetation classes and subclasses used in forest type classification in the Upper Potaro River Basin, Guyana. Parent classes are shown on left.....………………………………………………………………... 19 2 Error matrix for maximum likelihood supervised classification of Landsat 5 TM scene (path 232, row 056, WRS 2) acquired on 9 August 1989. Land cover classes for the Upper Potaro River Basin (Guyana) study site include: (1) Dicymbe corymbosa dominant and monodominant forest; (2) Micrandra dominant forest; and (3) mixed species forest lacking D. corymbosa ………...... 23 3 Error matrix for maximum likelihood supervised classification of Landsat 7 ETM+ scene (path 232, row 056, WRS 2) acquired on 16 October 1999. Land cover classes for the Upper Potaro River Basin (Guyana) study site include: (1) Dicymbe corymbosa dominant and monodominant forest; (2) Micrandra dominant forest; and (3) mixed species forest lacking D. corymbosa ……………………………………………………………..…….. 24
vii
LIST OF FIGURES
Figure Page 1 Study area showing position of Guyana in South America and the Pakaraima Montane region……………………………………………………………...….. 15 2 Thematic map showing predicted forest distribution for the Upper Potaro River Basin study site in western Guyana. Supervised classification output for 16 October 1999 Landsat 7 ETM+ scene is shown. Cloud covered areas were filled with classification output from the Landsat 5 TM image acquired on 9 August. Green areas represent Dicymbe corymbosa dominant and monodominant forest, blue depicts Micrandra dominant forest, and orange indicates mixed species forest lacking D. corymbosa .……………………...….. 25
viii
INTRODUCTION
Dominance of tropical rainforests by a single species of canopy tree ( i.e. monodominance) is a phenomenon of interest to tropical ecologists (Henkel 2003, Torti et al . 2001). Single species dominance may result from life history traits of the dominant tree species that alter the understory environment, promote self persistence and recruitment, and lead to clumped distributions ( e.g ., ectomycorrhizal habit, mast fruiting, limited dispersal, shade tolerance, reiteration; Connell & Lowman 1989, Hart et al. 1989,
Torti et al . 2001, Henkel et al. 2005a, Woolley et al. 2007). Monodominant forests allow researchers to study ecological processes which circumvent the typically high tree species diversity in tropical rainforests (Connell and Lowman, 1989, Hart 1990, Torti et al. 2001,
Leigh et al. 2004).
Monodominance has been documented in both the New and Old World tropics, but the phenomenon is not uniformly distributed across these regions. Forest monodominance by caesalpinioid legumes ( e.g. Gilbertiodendron dewevrei ) is extensive across portions of the Congo Basin (Torti et al. 2001). In lowland tropical Asia certain dipterocarps ( e.g . Drybalanops aromatica ) achieve monodominance (Whitmore 1984,
Richards 1996). Monodominance is less well known in the Neotropics, where the only documented examples of persistent dominance (ample regeneration of the dominant species, unrelated to edaphic influence, flooding or succession) involve the caesalpinioids
Peltogyne gracilipes in northern Brazil (Nascimento et al. 1997) and Dicymbe spp. in
Guyana (Zagt 1997, Henkel et al. 2002, Henkel 2003).
The ectomycorrhizal canopy tree Dicymbe corymbosa Spruce ex Benth. achieves dominance levels among the highest recorded for upland monodominant species in the
9 10
Neotropics. In the Pakaraima Mountains of western Guyana, D. corymbosa comprises 63 to 95 percent of the total basal area in stands ranging from one to several hectares
(Henkel 2003). In these stands, the species dominates all levels of forest strata due to reiteration and persistence of mature individuals as well as high recruitment of shade tolerant seedlings and saplings (Henkel 2003, Henkel et al. 2005a, Woolley et al. 2007).
Additionally, large amounts of organic matter that accumulate on the complex reiterated trunks and root mounds of D. corymbosa and their permeation by ectomycorrhizal rootlets may constitute a mineral nutrient recycling system that helps preclude establishment of non ectomycorrhizal tree species (Mayor & Henkel 2006, Woolley et al. 2007).
While plant diversity in D. corymbosa forests is greatly reduced from that of surrounding mixed species forests, fungal diversity may be positively affected. In local areas of the Upper Ireng and Upper Potaro River Basins under study, obligately ectomycorrhizal fungi largely appear to be spatially restricted to Dicymbe stands and absent from the mixed forest matrix (Henkel et al. 2002, 2006). The ectomycorrhizal fungal guild associated with Dicymbe at these sites currently numbers about 150 species, most of which are new to science ( e.g. Henkel 1999, 2001, Miller et al. 2002, Matheny et al. 2003, Henkel et al. 2005b, Fulgenzi et al. 2007 ). The discovery of the Dicymbe associated ectomycorrhizal fungal guild in Guyana effectively doubled the number of known Neotropical species of ectomycorrhizal fungi. The spatially limited, site specific distribution of ectomycorrhizal fungi in an otherwise non ectomycorrhizal forest matrix suggests Dicymbe forests may function as the primary, if not exclusive, habitats for these obligately symbiotic fungi in Guyana (Henkel et al. 2002).
11
Ground based observations of D. corymbosa thus far have been limited to a local scale; however, there is a need for wider distribution information (Fanshawe 1952,
Richards 1996, Henkel 2003). This knowledge would allow broader sampling of D. corymbosa stands, providing further insight into the phenomenon of persistent tropical monodominance. Distribution information would provide essential guidance on appropriate future conservation action for habitat acting as a major reservoir of
Neotropical ectomycorrhizal fungal diversity. Several factors suggest satellite remote sensing as an appropriate tool to examine this relatively unknown tropical forest system.
Remote sensing can be a cost effective solution to studying land cover in tropical regions where ground based data are difficult to obtain due to inaccessibility or high acquisition costs (Clark et al. 2005). Spectral reflectance data from remote optical sensors can be used to extract different types of information from target vegetation
(Cohen & Goward 2004 , Boyd & Danson, 2005). This technology has been applied to tropical forests to document anthropogenic and natural disturbance (Skole & Tucker
1993, Read et al . 2003, Souza et al . 2005), detect changes in phenology, type and extent of vegetation (Green et al. 1994, Bohlman et al .1998, Carreiras et al . 2006), investigate succession, community composition, and species diversity (Vieira et al. 2003, Gillespie et al. 2004, Salovaaraa et al. 2005), and create land cover maps (Mayaux et al .1998, Eva et al. 2004). Remote sensing techniques can be a powerful tool to generalize from plot scale, site specific studies typical of tropical field studies up to landscape, regional and global scales (Foody et al . 2003, Clark et al . 2005).
Remotely sensed data from Landsat satellites have been used to map forests in tropical regions, although difficulties in discriminating distinct forest types persist (Foody
12 et al. 2003, Cohen & Goward, 2004, Powell et al. 2004, Thenkabail et al . 2004,
Salovaaraa et al. 2005, Carreiras et al . 2006, Fuller 2006). In addition to its relatively low cost, Landsat imagery has a globally comprehensive data acquisition policy, centralized online searchable databases, and 30 meter spatial resolution across six optical bands (Fuller 2006). Landsat's spectral resolution is well suited for vegetation classification since it includes the near infrared (NIR) and shortwave infrared (SWIR) portions of the electromagnetic spectrum (NASA 2003). These bands are most effective in separating the spectra of tropical canopy tree species (Cochrane 2000, Clark et al.
2005). In order to identify a specific type of plant community in satellite imagery, the target vegetation must have a unique spectral signature within the context of its surroundings (Lillesand & Kiefer 2004; Clark et al. 2005, Zhang et al. 2006). This spectral signature is dependent on the properties of plant tissues, the physiognomic structure of the vegetation, and numerous other contributing factors (Lillesand & Kiefer
2004). Differentiation of vegetation type classes is generally controlled by spectral variation within classes (intra class) and spectral variation between classes (inter class)
(Clark et al. 2005, Zhang et al . 2006).
Detailed vegetation classification presents challenges in the tropics, where high species diversity, structural complexity, and cloud cover introduce complexity into radiance signals (Asner 2001, Zhang et al. 2006). However, structurally uniform and contiguous stands of tropical monodominant forest may be detectable with Landsat imagery. Dicymbe corymbosa stands in western Guyana have a high probability of classification success given their uniformity of crown structure, leaf composition, and
13 vertical stratification, in contrast to the surrounding, structurally complex and species diverse forest matrix (Henkel 2003, Woolley et al. 2007).
In this study, I examine the suitability of Landsat satellite imagery for mapping distribution of D. corymbosa monodominant forests in Guyana’s Upper Potaro River
Basin. Specifically, I asked: Can supervised image classification of Landsat 5 TM and
Landsat 7 ETM+ data allow me to distinguish monodominant Dicymbe stands from mixed forests and other forest types in the Pakaraima Mountain region? If successful, this approach could potentially be used to develop distribution estimates of ecologically important monodominant forests in other remote tropical regions.
STUDY SITE
Research was conducted in the central Pakaraima Mountains of western Guyana, an area hypothesized to be the geographical distribution center of D. corymbosa
(Fanshawe 1952, Henkel 2003, Figure 1). The central Pakaraimas region is blanketed with dense primary forest, predominantly of the seasonal evergreen Eschweilera –
Licania association with elevations ranging from 700 2200 meters (Fanshawe 1952).
Limited small–scale mining and slash and burn agriculture has taken place in the region; however, these activities have not caused visible changes in forest structure. Primary stands dominated by D. corymbosa are found within hilly, intermountain valleys of the adjacent Upper Potaro, Mazaruni, and Ireng River Basins (Myers 1936, Fanshawe 1952,
Henkel 2003, unpubl. data). These stands appear to vary in size from roughly one to several hectares, and have well defined boundaries with the surrounding, diverse, mixed– species forest matrix which lacks D. corymbosa (Fanshawe 1952, Henkel 2003, Henkel et al. 2005). Stands dominated by Micrandra glabra Schultes (Euphorbiaceae) and
Micrandra spruceana Schultes occur in extensive contiguous patches – on low lying, poorly drained sandy soils overtopping hardpan or sheetrock (Fanshawe 1952). Further details of the climate, geology, and soils of the region can be found in Henkel (2003),
Henkel et al . (2005) and Mayor & Henkel (2006).
14 15
Figure 1. Study area showing position of Guyana in South America and the Pakaraima Montane region.
MATERIALS AND METHODS
Satellite Data
Two Landsat satellite images (170 x 183 km) encompassing the study area were obtained from the Global Land Cover Facility of the University of Maryland
(http://glcf.umiacs.umd.edu/). Satellite data consisted of one Landsat 7 ETM+ scene
(path 232, row 056, WRS 2) acquired on 16 October 1999, and one Landsat 5 TM scene
(path 232, row 056, WRS 2) acquired on 9 August 1989. Both images were used in the forest classification process, and selected for low (<10 percent) cloud cover over the study area with the assumption that tree species distribution had not changed significantly in the interval to the present. Spectral resolution of the imagery is 0.45 to 2.35 m with seven bands spanning the blue to mid infrared portion of the electromagnetic spectrum
(NASA 2003). Bands 1 – 5 and 7 were included in the analysis. Spatial resolution (pixel size) of the imagery is 28.5 meters (NASA 2003). Images were orthorectified and co registered by EarthSat (Global Land Cover Facility 2007). Both scenes were projected in the Universal Transverse Mercator Projection, zone 20 N, using the WGS 1984 datum.
ERDAS Imagine Software version 9.1 was used for analysis of satellite imagery data
(Leica Geosystems 2006).
Preliminary Data Processing
Preliminary supervised classification of the Landsat images was used to develop an effective stratified sampling regime for wider scale field data collection and accuracy assessment. Supervised classification incorporates a priori knowledge from field sampling into the classification process (through the definition of training areas) and relates measured spectral reflectance properties directly to known vegetation cover
16 17 properties. For each pre determined category, training areas are identified on the image and delineated so their spectral properties can be examined. Classification of the remainder of the image is based on values defined via these training sets (Lillesand and
Kiefer, 2004). Preliminary supervised classification was carried out using datasets compiled from eight one hectare forest inventory plots of known structure and tree species composition (Henkel 2003, Henkel et al. 2005, Woolley et al. 2007) and 65 additional transect point surveys. This work was conducted by Dr. Terry Henkel
(Department of Biological Science, Humboldt State University) in June August 2006 within a five kilometer radius of a permanent base camp located along the Upper Potaro
River at 5° 18' 04.8" N; 59° 54' 40.4" W, at 710 750 meters elevation. Forest data were simplified for classification, and field sites were assigned to one of three classes: (1) D. corymbosa dominant (> 60 percent of stems ≥ 10 cm dbh and/or stand basal area) to monodominant (>80 percent of stems ≥ 10 cm dbh and/or stand basal area); (2)
Micrandra dominant (>60 percent stems ≥ 10 cm dbh and/or stand basal area); or (3) mixed species forest lacking D. corymbosa and Micrandra spp. Transition zones between forest types were excluded from the analysis. I utilized forest composition ground data, Global Positioning System (GPS) data, and topographic maps to digitize polygons around areas of known forest types in a Geographic Information System
(Environmental Systems Research Institute 2006).
These polygons were subsequently used as training areas for preliminary supervised classification of Landsat images. I used the output classification map to identify contiguous forest patches larger than five hectares over a broader area of the
Upper Potaro Basin as suitable sampling targets from which to collect representative field
18 data for subsequent analyses. The latitude and longitude of weighted internal centroids of selected target patches were calculated and then uploaded to a handheld GPS (Garmin
GPSMap76).
Field work for the secondary phase of D. corymbosa classification and accuracy assessment was conducted in Guyana during January 2007, within a twelve kilometer radius of a camp located approximately three kilometers south of the Upper Potaro River at 5° 17' 22.6" N; 59° 52' 23.9" W. This camp was approximately eight kilometers east of the base camp described above. My field team and I navigated via compass and GPS to patches of predicted target vegetation wherein I randomly chose a minimum of three locations at which to assess forest composition. In total, coordinates of 182 GPS survey points were recorded. Each point was located at least 60 meters into a contiguous forest type and 180 meters from other survey points. Species composition and relative abundance of canopy trees (~ ≥ 30 cm dbh) within a 30 meter quadrat were estimated and recorded by four tree spotters. Vegetation data for each location was input into a GIS database and, for classification purposes, assigned to one of eight predefined subtype
classes as defined in Table 1.
19
Table 1. Characteristics of vegetation classes and subclasses used in forest type classification in the Upper Potaro River Basin, Guyana. Parent classes are shown on left. Class Description Sub class Description
DCD D. corymbosa dominant
DCM D. corymbosa monodominant
a Dicymbe corymbosa dominant D. corymbosa dominant, to monodominant b, Dicymbe DCF DCDADC Dicymbe altsonii present to altsonii absent, present, or codominant codominant D. corymbosa dominant to monodominant, D. altsonii DCDACM present to contributing to monodominance
MGD Micrandra glabra dominant MF Micrandra sp. Dominated a Micrandra spruceana SWD dominant Mixed species riverine, MXR Non Dicymbe corymbosa lowland association MX containing mixed species forest MX Mixed species upland forest a ≥ 60 percent stems and/or basal area; b ≥ 80 percent stems and/or basal area
20
Classification Procedure
To improve vegetation classification, imagery areas unsuitable for target vegetation (non forest savanna, scrubland and water), clouds and shadows were excluded from consideration using preliminary unsupervised classification (30 classes) of each image (Tottrup 2004, ERDAS IMAGINE 2005). Unsupervised classification is a purely statistical method and incorporates no a priori knowledge of target characteristics into the classification process. Digital values from combinations of spectral bands are used to identify inherent groupings in the data (Lillesand & Kiefer 2004). Classes representing the aforementioned unsuitable categories were identified via visual interpretation, and these pixels were masked from subsequent classification procedures (ERDAS IMAGINE
2005).
In keeping with my objectives, I chose a final classification scheme identifying:
(1) D. corymbosa dominant and monodominant forest; (2) Micrandra dominant forest; and (3) mixed species forest lacking D. corymbosa . Several subclasses existed within each parent class and were retained in order to achieve maximum spectral separability between classes (Lillesand & Kiefer 2004). Subclasses were pooled post classification.
Forty percent of forest survey points from each subclass were randomly selected to train the supervised classification process; the remaining sixty percent were set aside for accuracy assessment. To accommodate survey quadrat size and GPS error, training points were buffered by 30 meters in the GIS and pixels within this buffer boundary were extracted to generate class specific spectral signatures (ERDAS IMAGINE 2005). A transformed divergence calculation was used to determine spectral separability between forest cover classes and as a guide to either aggregate or discard classes (Cayuela et al .
21
2006). A maximum likelihood supervised classification was subsequently performed on both Landsat images using spectral signatures of the retained classes. Output was thresholded at α = 0.05 to exclude pixels which did not fit given classes well (ERDAS
IMAGINE 2005).
Thematic Accuracy Assessment
The output classification map was compared with ground reference data to assess the accuracy of my remote sensing derived product (Foody 2002, Powell et al. 2004).
Sixty percent of the field survey points in each class were excluded from supervised classification training and set aside for this purpose. For both images survey points occurring within masked areas for a given date were not used. Classification accuracy was ultimately assessed using 134 independent reference points. To examine relationships between classification and reference data, I generated error matrixes for each classification map and calculated the following statistics: (1) errors of omission, (2) errors of commission, and (3) Khat , an estimate of the conditional Kappa coefficient expressing the proportional reduction in accuracy, when random chance agreement is removed (Congalton & Green 1999). “User’s accuracy”, the probability that a pixel classified on the image actually is that land cover type on the ground, was selected as the accuracy assessment parameter most appropriate to future use of this approach for field survey design (Congalton & Green 1999). Also in keeping with the purpose of the study, accuracy assessment focused on successful classification of single species dominant forest classes (Foody 2002).
22
RESULTS
For both the 1989 and 1999 images classification error matrixes show good agreement between the classified map and reference data for the D. corymbosa and
Micrandra forest classes (Tables 2 and 3). For the 1989 and 1999 images, the conditional
Kappa statistic values for user’s accuracy for the D. corymbosa class were 0.735 and
0.587 respectively, indicating the classifications were 73.5 and 58.7 percent correct accounting for chance agreement. The Micrandra forest class had the highest class accuracy, with conditional Kappa statistic values for user’s accuracy of 0.841 (for 1989) and 0.948 (for 1999). For both dates, the D. corymbosa class had lower errors of commission (10.2 percent for 1989; 19.3 percent for 1999) than errors of omission (41.3 percent for 1989; 23.6 percent for 1999). For the 1989 image, my classification correctly identified 58.7 percent of the D. corymbosa forest present, “producer’s accuracy”. The probability of a pixel labeled as D. corymbosa on the classified image actually being D. corymbosa in situ is 89.8 percent, - “user’s accuracy” (Congalton & Green 1999). For the 1999 image producer’s accuracy was 71.4 percent, and user’s accuracy 80.7 percent.
Low conditional Kappa statistic values for the mixed species forest class (0.221 for 1989;
0.347 for 1999) reflect high omission and commission errors for this class. The most common misclassification was D. corymbosa forest being misclassified as mixed species forest.
My classification output maps indicate D. corymbosa is locally prevalent across the study area; its stands are interspersed with patches of mixed species forest.
Micrandra dominated forest is present in more internally homogeneous swaths north of the Potaro River (Figure 2).
23
Table 2. Error matrix for maximum likelihood supervised classification of Landsat 5 TM scene (path 232, row 056, WRS 2) acquired on 9 August 1989. Land cover classes for the Upper Potaro River Basin (Guyana) study site include: (1) Dicymbe corymbosa dominant and monodominant forest; (2) Micrandra dominant forest; and (3) mixed species forest lacking D. corymbosa.
Reference data User’s Error of Classification data DCF MF MX Total accuracy commission Khat (%) (%) Dicymbe corymbosa 44 0 5 49 89.8 10.2 0.735 forest (DCF) Micrandra forest (MF) 2 20 1 23 87.0 13.0 0.841 Non DCF, mixed species 29 2 19 50 38.0 62.0 0.221 forest (MX) Total 75 22 25 122
Producer’s accuracy (%) 58.7 90.9 76.0
Error of omission (%) 41.3 9.1 24.0
24
Table 3. Error matrix for maximum likelihood supervised classification of Landsat 7 ETM+ scene (path 232, row 056, WRS 2) acquired on 16 October 1999. Land cover classes for the Upper Potaro River Basin (Guyana) study site include: (1) Dicymbe corymbosa dominant and monodominant forest; (2) Micrandra dominant forest; and (3) mixed species forest lacking D. corymbosa. Reference data User’s Error of Classification data DCF MF MX Total accuracy commission Khat (%) (%) Dicymbe corymbosa 42 1 9 52 80.7 19.3 0.587 forest (DCF) Micrandra forest (MF) 1 25 0 26 96.2 3.8 0.948 Non DCF, mixed species 12 1 12 25 48.0 52.0 0.347 forest (MX) Total 55 27 21 103
Producer’s accuracy (%) 71.4 92.6 57.1
Error of omission (%) 23.6 7.4 42.9
25
Figure 2. Thematic map showing predicted forest distribution for the Upper Potaro River Basin study site in western Guyana. Supervised classification output for 16 October 1999 Landsat 7 ETM+ scene is shown. Cloud covered areas were filled with classification output from the Landsat 5 TM image acquired on 9 August. Green areas represent Dicymbe corymbosa dominant and monodominant forest, blue depicts Micrandra dominant forest, and orange indicates mixed species forest lacking D. corymbosa .
DISCUSSION
Classification of Forest Types
Supervised classification of Landsat imagery performed well in distinguishing forest types dominated by a single tree species from mixed species forest. I achieved an accuracy of > 80 percent user’s accuracy for D. corymbosa (80.7, 89.9 percent) and
Micrandra (87.0, 96.2 percent) dominated stands. This accuracy level can be considered acceptable, since > 85 percent accuracy is often set as an upper target for less complex land cover classes (Foody 2002). The most common misclassification was D. corymbosa forest being incorrectly labeled as mixed species forest (Table 2, 3). This misclassification may arise from: 1) combining multiple types of non D. corymbosa, mixed species forest into two general categories (Table 1), creating high spectral variation within the class, and 2) relatively low spectral separability observed between D. corymbosa and mixed species forest classes. Spectral reflectance of monodominant forests is likely influenced by within stand uniformity of crown structure, leaf composition, and vertical stratification, allowing discrimination from structurally complex, species diverse forests (Henkel 2003). In addition, D. corymbosa ’s reiterative coppicing may play an important role in reducing canopy liana load, thus reducing spectral variation of within stand canopy reflectance in contrast to mixed species, liana rich forest (Richards 1996, Henkel 2003, Clark et al. 2005, Woolley et al. 2007).
Spectral separation between the Micrandra forest class and the D. corymbosa and mixed species forest classes may be accentuated by a high vegetation moisture content of
Micrandra ’s canopy and the presence of water beneath the forest (Mayaux et al. 2000).
Both of these conditions can strongly influence reflectance in short wave infrared
26 27
(SWIR) portion of the electromagnetic spectrum (Mayaux et al . 2000, Lillesand & Kiefer
2004).
Direct comparisons between my results and other published research is challenging because previous remote sensing studies in the tropics focused on mapping broad, general primary and secondary forest cover categories, with little fine scale forest type differentiation (Hansen et al . 2000, Mayaux et al. 2000, Vieira et al 2003, Eva et al .
2004, Stibig et al . 2004, Thenkabail et al . 2004, Joshi et al . 2006). Further complications arise when studies do not include detailed information on vegetation class definitions, sampling design, and individual class accuracy assessment (Foody 2002, Powell et al.
2004, Salovaara et al . 2005).
In some cases, I found that my Dicymbe and Micrandra forest class accuracies were comparable to those obtained for evergreen broadleaf tropical forests in South
America and India (Hansen 2000, Joshi et al. 2006). In cases where detailed forest cover discrimination was problematic, my monodominant forest class accuracies exceeded those of prior studies. These studies found various primary tropical canopy forest classes could not be separated from other cover classes, such as mixed deciduous, floodplain, semi evergreen, degraded primary, and late secondary successional forest, with multi spectral sensor data (Foody & Hill 1996, Mayaux 2000, Stibig et al . 2004, Thenkabail et al . 2004, Carreiras et al. 2006). However, in the Peruvian Amazon, Salovaara et al. 2005 used Landsat ETM+ and elevation data to differentiate among three floristically defined types of primary lowland tropical forest, with class accuracies ranging from 48.0 to 94.6 percent user’s accuracy. My study and that of Salovaara et al . (2005) both focused on classifying relatively structurally homogeneous, closed canopy primary tropical forest.
28
Their work differs from mine in that their forest classes are still somewhat broader and likely of higher tree species diversity and internal heterogeneity than my monodominant forest classes (Salovaara et al. 2005). Both Hill (1999) and Mayaux et al . (2000) found low diversity, closed canopy swamp forest could be separated from other forest classes by spectral reflectance.
Regional Distribution
Dicymbe corymbosa forest appears particularly widespread over a 40 km stretch south of the Upper Potaro River to the Upper Ireng River Basin (Fanshawe 1952,
Richards 1996, Henkel et al. 2005). My classification map indicates D. corymbosa monodominant forest occurs on lowland white sand drainages, as well as on ironstone hills with which it is often associated (Fanshawe 1952, Henkel 2003). Interestingly, for the D. corymbosa class, both of my classifications (1989 & 1999) had lower errors of commission (producer’s accuracy) than of omission (user’s accuracy). My classification identified only 58.7 (1989) to 71.3 (1999) percent of total D. corymbosa present; however, the area it classified as D. corymbosa has an 89.8 (1989) to 80.7 (1999) percent chance of actually being D. corymbosa in situ . This makes my classified D. corymbosa map and thus my estimate of the species’ distribution quite conservative in that when D. corymbosa was mapped, there is a high probability it is actually there. However, there is also a good chance (41.3 percent, 1989 image; 23.6 percent, 1999 image) some stands of
D. corymbosa were missed.
My D. corymbosa classification likely defines the core areas of ectomycorrhizal fungal habitat in the study area; however, ectomycorrhizal fungi may not necessarily be restricted to this distribution. In particular, ectomycorrhizal fungi may occur in specific
29 locales within the mixed forest matrix in association with the ectomycorrhizal tree
Dicymbe altsonii (Henkel et al. 2002). In my field surveys, D. altsonii was found to occur patchily within diverse mixed species forest and in some cases as a co dominant in
D. corymbosa stands. I could not separate D. altsonii from D. corymbosa or mixed species forest by spectral reflectance. This may be a result of: 1) my small D.altsonii containing forest sample size, 2) the complexity of and variability among D. altsonii forest types, or 3) the physiognomy and structural characteristics of D.altsonii , which does not dominate the forest’s vertical strata to the same degree as D. corymbosa (Zagt
1997, Henkel et al. 2002). Dicymbe altsonii may act as low density sources of ectomycorrhizal within anectotrophic mixed species forest and provide important ectomycorrhizal community connectivity between D. corymbosa stands (Henkel et al .
2002).
My classification map also shows large swaths of Micrandra dominated forest extending north from the Upper Potaro River through blackwater drainage systems.
There has been little work conducted on these forests, which may be dominated by either
M. glabra or M. spruceana occurring on shallow sandy soils (Fanshawe 1952). As examples of tropical “heath” forests and their attendant blackwater drainage systems these forests could serve as study systems for regional investigations of these ecologically unique communities (Janzen 1974).
My use of supervised classification of Landsat imagery to map the distribution of
D. corymbosa and Micrandra dominated stands is particularly promising given previous studies’ difficulties in differentiating among primary tropical forest subtypes (Foody &
Hill 1996, Mayaux 2000, Thenkabail et al. 2004). However, other studies suggest further
30 steps that could be taken to refine my classification. In particular, image segmentation, multi date analysis, and pre classification image smoothing have been shown to increase
Landsat classification accuracies for tropical forest classes (Hill 1999, Tottrup 2003).
Hyperspectral sensors, which give detailed reflectance information for over 50 narrow spectral bands (spanning the visible, near infrared, and shortwave infrared regions of the electromagnetic spectrum), show great promise for detailed mapping of tropical forest classes (Cochrane 2000; Thenkabail et al . 2004; Clark et al . 2005; Zhang et al . 2006).
When compared to Landsat ETM+ imagery, hyperspectral data analysis allowed separation of complex tropical moist forest classes with much higher accuracy: 96 percent vs. 42 percent (Thenkabail et al . 2004). However, the high expense of acquiring hyperspectral data, the lack of currently operating platforms, and the advanced data storage and processing capacity necessary for analysis indicate that the use of hyperspectral data is not a viable solution for most studies (Fuller 2006, Zhang et al .
2006). In particular, its limited geographic and temporal coverage is problematic to researchers working in remote tropical regions with abundant cloud cover (Fuller 2006).
Landsat data’s affordability, widespread availability, globally comprehensive acquisition policy, and historic time series make it the attractive choice for most tropical studies
(Cohen & Goward 2004, Fuller 2006).
Future Applications
My study is the first to focus on identifying species specific tropical monodominant forest using remotely sensed data. I found Landsat imagery suitable for mapping the distribution of D. corymbosa monodominant forests in the Upper Potaro
River Basin. This approach should be useful in identifying sites for forest ecological
31 studies and ectomycorrhizal fungal surveys. Additionally, with proper pre processing of
Landsat imagery (Bruce & Hilbert 2004) this technique shows potential to estimate the regional distributions of Dicymbe corymbosa forests, and help guide studies involving regional scale processes contributing to monodominance, such as mast fruiting (e.g.
Henkel et al. 2005). High classification accuracy for Micrandra dominated forest indicates that my approach offers a relatively straightforward, inexpensive means of gaining a better understanding of this understudied forest association. I suggest supervised classification of Landsat imagery (employing detailed training data) may be applicable to mapping other tropical monodominant species ( e.g. Gilbertiodendron dewevrei and Drybalanops aromatica ) in remote regions, such as Africa’s Congo Basin and lowland tropical Asia.
LITERATURE CITED
Asner, G. P. 2001. Cloud cover in Landsat observations of the Brazilian Amazon. International Journal of Remote Sensing 22: 3855– 3862. Bohlman, S., J. B. Adams, M. O. Smith, and D. L. Peterson. 1998. Seasonal foliage changes in the Eastern Amazon Basin detected from Landsat Thematic Mapper satellite images. Biotropica 30: 376–391.
Boyd, D. S. and F. M. Danson. 2005. Satellite remote sensing of forest resources: three decades of research development. Progress in Physical Geography 29: 1–26. Bruce, C. M. and D. W. Hilbert. 2004. Pre processing Methodology for Application to Landsat TM/ETM+ Imagery of the Wet Tropics. Cooperative Research Centre for Tropical Rainforest Ecology and Management. Rainforest CRC, Cairns, Australia. Carreiras, J. B., Pereira, J., Campagnolo , M. L, and Y. E. Shimabukuro. 2006. Assessing the extent of agriculture/pasture and secondary succession forest in the Brazilian Legal Amazon using SPOT VEGETATION data. Remote Sensing of Environment 101: 283 298. Cayuela, L., D. J. Golicher, and J. M. Rey Benayas. 2006. The Extent, Distribution, and Fragmentation of Vanishing Montane Cloud Forest in the Highlands of Chiapas, Mexico. Biotropica 38: 544–554. Clark, M. L., Roberts, D. A., and D.B Clark. 2005. Hyperspectral discrimination of tropical rain forest tree species at leaf to crown scales. Remote Sensing of Environment 96: 375 398. Cochrane, M. A. 2000. Using vegetation reflectance variability for species level classification of hyperspectral data. International Journal of Remote Sensing 21: 2075−2087. Cohen, W. and S. N. Goward. 2004. Landsat’s role in ecological applications of remote sensing. Bioscience 54: 535 545. Congalton, R., and K. Green. 1999. Assessing the Accuracy of Remotely Sensed Data: Principles and Practices. CRC/Lewis Press, Boca Raton, Florida. Connell, J. H., and M. D. Lowman. 1989. Low diversity tropical rainforests: some possible mechanisms for their existence. American Naturalist 134: 88 119. Environmental Systems Research Institute. 2006. ArcGIS 9: Using ArcGIS Desktop. ESRI Press, Redlands, California.
32 33
ERDAS IMAGINE. 2005. Erdas Imagine Field Guide. ERDAS Inc., Atlanta, Georgia. Eva, H.D., A.S. Belward, E.E. DeMiranda, C.M. DiBella, V. Gond, O. Huber, S. Jones, M. Sgrenzaroli, and S. Fritz. 2004. A land cover map of South America. Global Change Biology 10: 731–744. Fanshawe, D. B. 1952. The vegetation of British Guiana: a preliminary review. Institute paper No. 39, Imperial Forestry Institute, Oxford,UK. Foody, G.M. 2002. Status of land cover classification accuracy assessment. Remote Sensing of Environment 80: 185 201. Foody, G. M., Boyd, D. S., and M. E. Cutler. 2003. Predictive relations of tropical forest biomass from Landsat TM data and their transferability between regions. Remote Sensing of Environment 85: 463−474. Foody, G.M. and R.A. Hill. 1996. Classification of tropical forest classes from Landsat TM data. International Journal of Remote Sensing 17: 2353 2367. Fulgenzi, T.D., Henkel, T.W., and R. Halling. 2007. Tylopilus orsonianus sp. nov. and Tylopilus eximius from Guyana. Mycologia (in press). Fuller, D. O. 2006. Tropical forest monitoring and remote sensing: A new era of transparency in forest governance? Singapore Journal of Tropical Geography 27: 15–29. Gillespie, T. W., Brock, J., and C. W. Wright. 2004. Prospects for quantifying structure, floristic composition and species richness of tropical forests. International Journal of Remote Sensing 25: 707–715. Global Land Cover Facility. 2007. GLCF Landsat Technical Guide: http://glcf.umiacs.umd.edu/data/guide/technical/landsat.shtml. Green, K., D. Kempka, and L. Lackey. 1994. Using remote sensing to detect and monitor land cover and land use change. Photogrammetric Engineering and Remote Sensing 60: 331 337. Hansen, M.C., R.S. DeFries, and J.G. Townshend. 2000. Global land classification at 1 km spatial resolution using a classification tree approach. International Journal of Remote Sensing 21 : 1331–64. Hart, T. B. 1990. Monospecific dominance in tropical rain forests. Trends in Ecology and Evolution 5: 6 11.
34
Hart, T. B., J. A. Hart, and P. G. Murphy. 1989. Monodominant and species rich forests of the humid tropics: causes for their occurrence. American Naturalist 133: 613 633. Henkel, T.W. 1999. New taxa and distribution records for Tylopilus from Dicymbe forests of Guyana. Mycologia 91: 655 665. Henkel, T.W. 2001. Tylopilus pakaraimensis sp. nov., a new bolete in Tylopilus section Potamogetones from Guyana. Mycotaxon 78: 105 114. Henkel, T. W. 2003. Monodominance in the ectomycorrhizal Dicymbe corymbosa (Caesalpiniaceae) from Guyana. Journal of Tropical Ecology 19: 417 437.
Henkel, T.W., J. R. Mayor, and L. P. Woolley. 2005a. Mast fruiting and seedling survival of the ectomycorrhizal, monodominant Dicymbe corymbosa (Caesalpiniaceae) in Guyana. New Phytologist 167: 543 556. Henkel, T.W., R. Meszaros, M.C. Aime, and A. Kennedy. 2005b. New species of Clavulina from the Pakaraima Mountains of Guyana. Mycological Progress 4: 342 350. Henkel, T.W., T.Y. James, S.L. Miller, M.C. Aime, and O.K. Miller, Jr. 2006. The mycorrhizal status of Pseudotulostoma volvata (Elaphomycetaceae, Eurotiales, Ascomycota). Mycorrhiza 16: 241 244. Henkel, T.W., J. Terborgh, and R. Vilgalys. 2002. Ectomycorrhizal fungi and their leguminous hosts in the Pakaraima Mountains of Guyana. Mycological Research 106: 515 531. Hill, R. A. 1999. Image segmentation for humid tropical forest classification in Landsat TM data. International Journal of Remote Sensing 20: 1039 – 1044. Janzen, D. 1974. Tropical blackwater rivers, animals, and mast fruiting. Biotropica 6: 69 103. Joshi, P.K., P.S. Roy, S. Singh, S. Agrawal, and D. Yadav. 2006. Vegetation cover mapping in India using multi temporal IRS Wide Field Sensor (WiFS) data. Remote Sensing of Environment 103: 190 202. Leigh, E.G, P. Davidar, C.W. Dick, J.P. Puyravaud, J. Terborgh, H. ter Steege, and S.J. Wright. 2004. Why do some tropical forests have so many species of trees? Biotropica 36: 447 473. Lillesand, T. M., and R. W. Kiefer. 2004. Remote sensing and image interpretation, 4th edition. John Wiley, New York, New York.
35
Matheny, P. B., M.C. Aime, M. C., and T.W. Henkel. 2003. New species of Inocybe from Dicymbe forests of Guyana. Mycological Research 107: 495 505. Mayaux, P., F. Achard, and J.P. Malingreau. 1998. Global tropical forest measurements derivedfrom course resolution satellite imagery: a comparison with other approaches. Environmental Conservation 25: 37 52. Mayaux, P., G. De Grandi, and J.P. Malingreau. 2000. Central African forest cover revisited: a multisatellite analysis. Remote Sensing of the Environment 71: 183 196. Miller S.L., M.C. Aime, and T.W. Henkel. 2002. The Russulaceae of the Pakaraima Mountains of Guyana. New species of pleurotoid Lactarius . Mycologia 94: 545 553. Mayor, J. R., and T. W. Henkel. 2006. Do ectomycorrhizas alter leaf litter decomposition in monodominant tropical forests of Guyana? New Phytologist 169: 579 588. Myers, J., G. 1936. Savanna and forest vegetation of the Interior Guiana Plateau. Journal of Ecology 24: 162 184. Nascimento, M. T. and D. M. Villela. 1997. The forest structure, floristic composition and soils of an Amazonian monodominant forest. Edinburgh Journal of Botany 54: 1 38. National Aeronautics and Space Administration. 2003. Landsat 7 Science Data Users Handbook. Website: http://ltpwww.gsfc.nasa.gov/IAS/handbook/handbook_toc.html. Powell, R. L., Matzke, N., de Souza Jr., C., Clark, M., Numata, I., Hess, L. L., and D. A. Roberts. 2004. Sources of error in accuracy assessment of thematic land cover maps in the Brazilian Amazon. Remote Sensing of Environment 90: 221–234. Read, J. M, Clark, D. B. Venticinque, E. M And M. P. Moreira. 2003. Application of merged 1 m and 4 m resolution satellite data to research and management in tropical forests. Journal of Applied Ecology 40: 592–600. Richards, P. W., 1996. The Tropical Rainforest. An ecological study. 2nd edn. Cambridge University Press, Cambridge. Salovaaraa, J. K., Thesslerb, S., Malikc, R. N., and H. Tuomisto. 2005. Classification of Amazonian primary rain forest vegetation using Landsat ETM+ satellite imagery Remote Sensing of Environment 97: 39 51. Skole, D. and C. Tucker. 1993. Tropical deforestation and habitat fragmentation in the Amazon: Satellite data from 1978 to 1998. Science 260: 1905–1910.
36
Souza, C. M., Roberts, D. A., and M. A. Cochrane. 2005. Combining spectral and spatial information to map canopy damage from selective logging and forest fires. Sensing of Environment 98: 329 343. Stibig, H.J., F. Achard, and S. F ritz. 2004. A new forest cover map of continental southeast asia derived from SPOT VEGETATION satellite imagery. Applied Vegetation Science 7: 153 162. Thenkabail, P. S., Enclona, E.A., and M. S. Ashton. 2004. Hyperion, IKONOS, ALI, and ETM+ sensors in the study of African rain forests. Remote Sensing of Environment 90: 23 43. Torti, S. D., P. D. Coley, and T. A. Kursar. 2001. Causes and consequences of monodominance in tropical lowland forests. American Naturalist 157: 141 153. Tottrup, C. 2004. Improving tropical forest mapping using multi date Landsat TM data and pre classification image smoothing. International Journal of Remote Sensing 25: 717 730. Vieira, I. C. G., de Almeida, A. S., Davidson, E. A., Stone, T. A., de Carvalho, C. R., and J. B. Guerrero. 2003. Classifying successional forests using Landsat spectral properties and ecological characteristics in Eastern Amazonia. Remote Sensing of Environment 87: 470 481. Whitmore, T. C. 1984. Tropical Rainforests of the Far East. 2nd edn. Clarendon Press, Oxford. Woolley, L.P., T.W. Henkel, and S.C. Sillett. 2007. Reiteration in the monodominant tropical tree Dicymbe corymbosa (Caesalpiniaceae) and its potential adaptive significance. Biotropica (in press). Zagt, R. J. 1997 . Stem sprouting of Dicymbe altsonii in the tropical rainforest of Guyana: impact of soil type and potential for regeneration. Pages 55 72 in R. J. Zagt, editor. Tree demography in the tropical rainforest of Guyana. Tropenbos Guyana Programme, Georgetown, Guyana. Zhang, J., B. Rivard, A. Sánchez Azofeifa, and K. Castro Esau. 2006. Intra and inter class spectral variability of tropical tree species at La Selva, Costa Rica: Implications for species identification using HYDICE imagery. Remote Sensing of Environment 105: 129–141.