Assessing aboveground tropical forest biomass using Google Earth canopy images

Transcription

1 Ecological Applications, 22(3), 2012, pp Ó 2012 by the Ecological Society of America Assessing aboveground tropical forest biomass using Google Earth canopy images PIERRE PLOTON, 1,2 RAPHAËL PE LISSIER, 1,3,5 CHRISTOPHE PROISY, 3 THE O FLAVENOT, 1 NICOLAS BARBIER, 3 S. N. RAI, 4,6 AND PIERRE COUTERON 3 1 De partement d Ecologie, Institut Français de Pondiche ry, UMIFRE MAEE-CNRS 21, Puducherry India 2 IRD, Institut de Recherche pour le De veloppement, UMR AMAP, University of Yaounde I, Yaounde, Cameroon 3 IRD, UMR AMAP, F Montpellier, France Maha Gauri Aptt., MLA Layout, RMV II Stage, Bangalore India Abstract. Reducing Emissions from Deforestation and Forest Degradation (REDD) in efforts to combat climate change requires participating countries to periodically assess their forest resources on a national scale. Such a process is particularly challenging in the tropics because of technical difficulties related to large aboveground forest biomass stocks, restricted availability of affordable, appropriate remote-sensing images, and a lack of accurate forest inventory data. In this paper, we apply the Fourier-based FOTO method of canopy texture analysis to Google Earth s very-high-resolution images of the wet evergreen forests in the Western Ghats of India in order to (1) assess the predictive power of the method on aboveground biomass of tropical forests, (2) test the merits of free Google Earth images relative to their native commercial IKONOS counterparts and (3) highlight further research needs for affordable, accurate regional aboveground biomass estimations. We used the FOTO method to ordinate Fourier spectra of 1436 square canopy images ( m) with respect to a canopy grain texture gradient (i.e., a combination of size distribution and spatial pattern of tree crowns), benchmarked against virtual canopy scenes simulated from a set of known forest structure parameters and a 3-D light interception model. We then used 15 1-ha ground plots to demonstrate that both texture gradients provided by Google Earth and IKONOS images strongly correlated with field-observed stand structure parameters such as the density of large trees, total basal area, and aboveground biomass estimated from a regional allometric model. Our results highlight the great potential of the FOTO method applied to Google Earth data for biomass retrieval because the texture biomass relationship is only subject to 15% relative error, on average, and does not show obvious saturation trends at large biomass values. We also provide the first reliable map of tropical forest aboveground biomass predicted from free Google Earth images. Key words: aboveground biomass; canopy texture; forest structure; Fourier spectra; Google Earth; tree biomass allometry; very-high-resolution images; Western Ghats of India. INTRODUCTION Deforestation and forest degradation have been shown to be the second most important source of anthropogenic carbon emissions, accounting for 20 25% of the total (IPCC 2007), although this figure has recently been revised downward to 10 15% (van der Werf et al. 2009). It is nevertheless a consensus that reducing carbon emissions from forest ecosystems (the UNFCCC REDD program: Reducing Emissions from Deforestation and Forest Degradation), particularly in the tropics where almost all of the emissions occur Manuscript received 2 September 2011; revised 1 November 2011; accepted 7 December Corresponding Editor: V. C. Radeloff. 5 Corresponding author. Present address: IRD, UMR AMAP, TA A51/PS2, Montpellier Cedex 05, France. Raphael.Pelissier@ird.fr 6 Formerly Principal Chief Conservator of Forest, Karnataka Forest Department, Bangalore, India (now retired). 993 (Houghton 2005), would be a cost-effective means to mitigate climate change. Although REDD is likely to be part of the future post-kyoto protocol (UNFCCC 2009), its implementation still faces a host of technical challenges. One basic requirement for practical application of the REDD mechanism is our technical ability to accurately assess forest carbon stock variations induced by deforestation and degradation processes. Such an assessment is typically achieved by combining spatially limited ground measurements of forest stand structure (e.g., trunk diameter distribution) with forest cover types extensively mapped from remote-sensing data (Maniatis and Mollicone 2010). Although deforestation stricto sensu (i.e., loss of forest cover) is fairly easy to measure and map using a variety of image types and methods (e.g., Hansen et al. 2008), forest degradation is far more difficult to monitor and is still hindered by technical limits in the tropics (DeFries et al. 2007). In particular, space-borne optical and radar signals of medium to high

2 994 PIERRE PLOTON ET AL. Ecological Applications Vol. 22, No. 3 spatial resolution, widely used over the last few decades, have been shown to saturate beyond quite low values of tropical forest biomass (; Mg/ha; Imhoff 1995, Mougin et al. 1999) or leaf area index (Huete et al. 2002, Foody 2003). Airborne systems such as lidar (Drake et al. 2002, St-Onge et al. 2008), although potentially powerful throughout the entire biomass range of tropical forests, are very expensive to operate, which limits their systematic use for large-scale assessments (but see Asner et al. 2010). The potential of the recent generation of very-high-resolution (VHR, i.e., with a pixel size ;1 m 2 ) optical data provided by IKONOS or QUICKBIRD satellites has not been fully investigated with regard to the study of tropical forest structure. Several attempts analyzing radiometric intensity (e.g., Thenkabail et al. 2004) or alternative approaches based on the delineation of individual crowns either from visual interpretation (e.g., Asner et al. 2002, Read et al. 2003) or automated methods (e.g., Broadbent et al. 2008), suggest that exploiting the geometrical properties of VHR canopy images could be a promising research avenue (e.g., Wulder et al. 1998, Frazer et al. 2005). Although crown delineation methods proved to be still biased for tropical forest applications, especially when tree crowns are small (not much larger than the pixel size) and/or jointed (Zhou et al. 2010), encouraging results have indeed been recently obtained from texturebased methods, notably on the basis of two-dimensional power spectral analysis by Fourier transform (e.g., Couteron et al. 2005, Proisy et al. 2007). Implementing REDD will require the monitoring of carbon stocks of potentially poorly known, large-biomass tropical forests (Defries et al. 2007) on regional to national scales. It calls for simple extrapolation methods able to produce reliable results over extensive areas at affordable prices (Maniatis and Mollicone 2010), for which remotesensing-based maps have undeniable potential. However, the acquisition cost of commercial VHR images at the REDD scale is a critical issue in many regions of the world. A common strategy, therefore, consists in extrapolating the results provided by space- or airborne-vhr data to a cheaper medium- to high-resolution coverage of the study area. Another possible approach arises from the development of virtual globe interfaces such as Google Earth (GE) that nowadays possess extensive coverage of freely available VHR images, at least for some regions of the tropics. Although GE data are of slightly lower quality than native-resolution genuine commercial images, they have proven sufficient to derive consistent characterizations of forest canopy texture (Barbier et al. 2009). In this paper, we used a promising texture-based analysis method, namely Fourier Textural Ordination (FOTO; Couteron 2002), to predict stand structure parameters, including aboveground biomass, from GE and IKONOS satellite images covering several thousand hectares of wet evergreen tropical forests in the Western Ghats (WG) of India. The model was calibrated using artificially generated canopy scenes from controlled virtual 3-D (three-dimensional) forest stands and a light interception model, as well as reliable estimates of the aboveground biomass of 15 1-ha field plots obtained from a regional allometric model built by revisiting a unique destructive data set (Rai 1981). Finally, we used the canopy texture-based model to produce what is, to the best of our knowledge, the first map of tropical forest biomass predicted from free GE images. MATERIAL Study site Our study site is a 30-km 2 area surrounding Uppangala Permanent Sample Plots (UPSP; N, E) at the Pushpagiri Wildlife Sanctuary in the Western Ghats of India (Pascal and Pe lissier 1996, Pélissier et al. 2011). It is located on the west-facing escarpment of the WG mountain range with a topography that ranges from 200 to 1000 m above sea level from the foothills to the crest and shows large variations due to the presence of frequent thalwegs that drain an annual rainfall in excess of 5000 mm. This uneven and barely accessible terrain hosts one of the last wellpreserved wet evergreen forests of the WG (Pascal 1988). The area is part of a forest reserve that was only partially subjected to light selective logging (,10 trees/ ha, unique rotation) in the 1970s 1980s (Loffeier 1989). Although logging effects are no longer substantial on current forest stand structure (Pélissier et al. 1998), degraded semievergreen forest stages are patchily distributed throughout the stand, caused by an intense wildfire disturbance that occurred about 25 years ago (Loffeier 1989), thereby providing a gradient of forest degradation levels. We extracted a m spatial resolution satellite image of the study area from GE interface by setting image size to pixels and user view to 4.15 km altitude. The RGB true color composite bands of this image were then averaged in a single grayscale layer. We also, for comparison, purchased the panchromatic image channel ( lm wavelength band, 1-m 2 spatial resolution) of the native IKONOS 2 image corresponding to that used by GE, acquired in January 2002 and delivered by Geoeye. No radiometric or geometric transformation was performed on the IKO- NOS image. For the sake of the FOTO analysis, both images were subsequently divided into m contiguous unit windows once non-forested areas (e.g., rubber plantations, bare soil) had been masked. Control data Three types of data served as controls for the study. First, to benchmark the discriminative ability of the FOTO method with respect to canopy grain, we used 10 artificially generated canopy scenes simulated on the same principles as in Barbier et al. (2009). We generated m virtual 3-D forest stands using a simple representation of trees with an ellipsoidal crown shape. The trees were assembled in two distinct layers, a canopy

3 April 2012 BIOMASS ASSESSMENT FROM CANOPY IMAGES 995 and an understory layer, with independently varying mean densities. In each layer, the trees were located according to a lattice-based regular pattern, i.e., by randomly locating individuals within radii centered on each node of a lattice whose internode distance is specified with respect to the desired mean layer density. In each layer, tree diameter at breast height (dbh), tree height, and crown diameter were simulated from distinct normal distributions (i.e., distinct means and standard deviations) with means linked by general allometric relationships for tropical forests as derived from the literature (for more details, see Barbier et al. 2009). A discrete anisotropic radiative transfer model (DART v.5; Gastellu-Etchegorry et al. 2004) was then used to simulate light interception by tree crowns in the 3-D virtual stands. Simulation parameters were set to produce canopy images in the visible domain with a m spatial resolution and under illumination conditions close to the acquisition conditions of the IKONOS native image (satellite azimuth angle ¼ 128, sun elevation angle ¼ 458, and relative sunsensor angle ¼ 1808). We thus obtained a set of realistic canopy images of virtual forest stands (see Fig. 1) with a modal crown diameter of the dominant layer designed to range from 7.5 to 25 m (for details about the ranges of tree morphological parameters used for each scene, see Appendix A: Table A1). Second, a regional allometric biomass model employed to estimate plot-level aboveground biomass (AGB) from individual tree diameter records was calibrated using the only data set of harvested trees directly weighted in the wet evergreen forests of the WG, by Rai (1981), and subsequently used in Rai and Proctor (1986) and Chave et al. (2005). The original data, published in Rai (1981), are reported as an electronic Supplement to this paper. They provide complete biomass data (i.e., dry mass of bole, branches, twigs, and leaves) for 189 trees of the 12 dominant species in Chakra wet evergreen forest, located about 100 km north of UPSP, in a structurally comparable forest type (Pascal 1988). Thirdly, we laid out 15 1-ha ground plots during two field campaigns in 2009 and 2010, with the aim of covering the entire canopy grain gradient observed in the image. We attempted to target only accessible zones possessing a homogeneous canopy texture in order to buffer uncertainty in plot geolocation. Each plot was a m square corrected for slope angle and georeferenced using a Trimble Juno SB GPS device (Trimble, Sunnyvale, California, USA). Plots from the 2010 field campaign were all oriented toward the cardinal points (N S/E W), allowing us to extract the corresponding m unit window directly from the satellite image based on plot center geolocation. Plots from the 2009 campaign had a random orientation, obliging us to average their textures from four m unit windows, half-overlapping in both horizontal and vertical directions (i.e., with an error margin of 62.5 m in both directions). Each ground plot was sampled for dbh (diameter at breast height or above the buttresses if any) of all trees with dbh. 10 cm. Four stand structural characteristics were computed from field data, for both the set of trees. 10 cm dbh (proxy for total stand) and the set of trees. 30 cm dbh (proxy for canopy trees): density (D), mean quadratic tree diameter (DQ or diameter of the tree of mean basal area), basal area (BA), and aboveground biomass (AGB) as estimated from the regional allometric model derived from Rai s (1981) data. METHODS Fourier textural ordination of canopy images We applied the FOTO method (FOurier Textural Ordination) in line with the procedure presented in Proisy et al. (2007) and using routines developed in the MatLab environment (MathWorks 2002). The satellite image of the study area was first divided into m contiguous unit windows, a size set to include at least five repetitions of the largest tree crown diameter, estimated to be ;25 m in UPSP (Robert 2003). Each unit window is a square matrix representing the spatial arrangement of the pixels spectral radiance (IKONOS: panchromatic; GE: grayscale), which can be transposed into the frequency domain by the two-dimensional Fast Fourier Transform (2D-FFT) algorithm (Ripley 1981). Detailed descriptions of Fourier spectral analysis (e.g., Ripley 1981) and its application to digital images (e.g., Couteron 2002) have been published elsewhere. Only its main points are outlined hereafter. The squared amplitude of the 2D-FFT yields a 2-D periodogram, which represents an apportionment of the variance of the pixels spectral radiance (or image variance) among spatial frequency bins in all possible planar directions within the geographic space defined by the digital image. Averaging the periodogram across all directions provides a radial- or r-spectrum that only extracts the scale information (directional information is neglected), efficiently quantifying the coarseness-related textural properties of each canopy window (Couteron et al. 2005, Proisy et al. 2007, Barbier et al. 2009). Each r-spectrum represents the frequency distribution of wavenumbers (or Fourier harmonic spatial frequencies), r, corresponding to the number of times a reference sine/cosine pattern repeats itself within the canopy grain of a given 125-m sided unit window. When applied separately to all unit windows, these steps result in a matrix of r-spectra, with as rows the individual canopy windows, and as columns the first 29 Fourier harmonic spatial frequencies, r, after discarding the two largest frequencies that only reflect macroscale patterns unrelated to canopy grain texture. To compare r-spectra derived from different window sizes or spatial resolutions, we can express spatial frequencies in cycles per km as f ¼ 1000rN 1 DS 1 (with DS being the pixel size in meters and N being the window size in pixels), giving the number of times an object repeats within 1 km, or use pattern sizes (wavelengths) in meters as k¼1000/f (typically the apparent diameter of emergent tree

4 996 PIERRE PLOTON ET AL. Ecological Applications Vol. 22, No. 3 FIG. 1. PCA on Fourier r-spectra (FOTO method) of 1436 square Google Earth canopy windows ( m) of a wet evergreen forest in the Western Ghats of India. Upper-left: histogram of eigenvalues giving the percentage of variance explained by each PCA axis in sequence. Upper-right: correlation circle between spatial frequencies (in cycles/km; values at ends of arrows) and PCA axes (selection of spatial frequencies above the 10th for legibility). Bottom: first factorial plane with black dots indicating the location of the virtual DART scenes with five samples (A E) shown for illustration. crowns for homogeneous canopy images; Barbier et al. 2009). Column-wise standardization is performed on the r- spectra matrix, which is then submitted to a principal component analysis (standardized PCA) with unit windows considered as observations characterized by the way their grayscale variance is broken down among the successive spatial frequencies, seen as quantitative variables. The latter usually end up positioned in a natural order in the first PCA plane (i.e., clockwise or counterclockwise distribution of the frequencies from the lowest to the highest; e.g., Couteron et al. 2005), and the cloud of unit windows can thus be interpreted in terms of canopy grain variation through changes in the relative contribution of spatial frequencies across the r-spectrum. Relating canopy texture to ground estimations of aboveground biomass A critical step known to be a major potential source of error in large-scale biomass assessments (e.g., Chave et al. 2004) is the development of an allometric model to

5 April 2012 BIOMASS ASSESSMENT FROM CANOPY IMAGES 997 predict tree AGB from nondestructive field measurements. We therefore used the Rai (1981) data set to fit a regional allometric model relating tree AGB to tree dbh. Although both tree height and wood specific gravity are available in Rai s data set, we only included dbh as a predictive variable in our regional allometric model so as to mimic minimal forest inventory data that are the most widely available in tropical regions. Moreover, wood specific gravity values have not been determined for all of the species present at our UPSP study site. After comparing the mathematical form of several commonly used allometric models (power, first- to thirdorder polynomial, with log-transformation, with absolute dbh or relative dbh at 10% height; results not shown), we selected a simple log-transformed power model with intercept set to 0: ln( dagb) ¼ ln(dbh); r 2 ¼ 0.998; RSE ¼ (where RSE is the residual standard error of the model). This model performed better than that proposed in Rai and Proctor (1986) when fitted with an intercept to the same data (r 2 ¼ 0.92), and also better than the models proposed in Chave et al. (2005) based on a worldwide data set and including wood specific gravity as a predictive variable in addition to dbh (r 2 ¼0.957 and 0.996, RSE ¼ and 0.357, for all forest types and moist forests, respectively, Rai s data being included in both cases). Although the moist forest pantropical model proposed by Chave et al. (2005) resulted in a lower relative prediction error, 12% vs. 23% for our model, we considered that these results validated our regional model with only dbh as a predictive variable. We then used this allometric model to compute AGB values for the 15 1-ha ground plots sampled for tree dbh at the study site (see Appendix A: Table A2). Because we used a log-transformed model that is known to introduce a 10 20% bias in AGB predictions, we multiplied the estimates by a correction factor (CF) calculated as CF ¼ exp(0.5 RSE 2 ) ¼ (see Brown et al. 1989, Chave et al. 2005). Finally, AGB estimates and forest structure parameters in the sampled plots were related to FOTO canopy grain indices by multiple regressions using unit window ordination scores along the main PCA axes as predictive variables. RESULTS Canopy grain texture analysis A preliminary visual assessment of the canopy windows extracted from GE images resulted in us excluded 108 unit windows (4.6% of the total set) that were particularly marked by topography-induced macro-heterogeneity (such as a slope alternation on both sides of a ridge line or a deep thalweg). We then submitted the remaining 1436 r-spectra to a standardized PCA. This yielded two prominent axes that synthesized more than 48% of the total variability of the data matrix (Fig. 1). The correlation circle, showing the relationships between the spatial frequencies and the PCA axes, gave the lowest values ( f 40 cycles/km) or equivalently largest pattern sizes (k 25 m) on the negative side of axis 1, and the highest spatial frequencies ( f 144 cycles/km) or smallest pattern sizes (k 7 m) on the opposite side, whereas intermediate spatial frequencies were correlated with the positive side of the second axis. We used these two axes to ordinate GE windows according to their textural properties, as illustrated by the positions of the virtual DART scenes in the PCA plane. Unit windows were sorted along a gradient of coarse (i.e., few large apparent crowns) to fine (i.e., numerous small apparent crowns) textures along a diagonal in the PCA plane. The IKONOS image provided equivalent results with a very similar percentage of variability (45%) explained by the first PCA plane and a comparable cloud shape (see Appendix B: Fig. B1), although the main gradient is closer to the first axis. A direct comparison of window coordinates against PCA axes revealed similarities between the first axes of the IKONOS vs. GE ordinations (Pearson r ¼ 0.68, P 0.001) and, to a lesser extent, between the second axes (Pearson r ¼ 0.25, P 0.001). Predicting stand structure parameters from canopy texture We then used multiple linear regressions of PCA scores for the DART, GE, and IKONOS canopy images on axes 1 and 2 (referred to as canopy texture indices in the following) to predict stand structure parameters (Table 1). Canopy texture indices of DART scenes show very strong relationships with all structural parameters. This benchmarking shows that the FOTO texture gradient covers realistic ranges of structure values and provides a consistent canopy grain classification (mean apparent crown diameter ranges from 7.5 to 25 m; R 2 ¼ 0.93 and 0.96 with relative RMSE ¼ 7.8% and 5.0% for GE- and IKONOS-derived canopy texture indices, respectively). When the model was applied to GE and IKONOS canopy windows corresponding to our 15 1-ha ground plots, it showed that canopy texture index, in both cases, is a good predictor of the density of the largest trees (R 2 ¼ 0.77 for both GE and IKONOS), but that it does not significantly relate to total tree density. Stand basal area (BA) and AGB estimates were also closely related to canopy texture indices, and these relationships remained fairly stable regardless of whether all trees were considered (R 2 ¼ 0.74 and 0.78 for both parameters with GE and IKONOS, respectively) or only the largest trees (R 2 ¼ 0.72 and for both parameters with GE and IKONOS, respectively). The relationship with mean quadratic tree diameter (DQ) was weaker but significant (R 2 ¼ 0.69 and 0.66 for GE and IKONOS, respectively), and weakened further when only the largest trees were considered (R 2 ¼ 0.39 and 0.54 for GE and IKONOS, respectively). FOTO-derived vs. dbh-derived AGB estimations With an average of,15% relative error on estimates, texture-based indices demonstrated a powerful ability to

6 998 PIERRE PLOTON ET AL. Ecological Applications Vol. 22, No. 3 Regression results between forest stand parameters (virtual DART scenes and field plots for trees.10 or.30 cm dbh) and FOTO textural indices derived from Google Earth and IKONOS canopy images for a wet evergreen forest in the Western Ghats of India. TABLE 1. Parameter Virtual vs. field plot Google Earth IKONOS R 2 adj. rrmse (%) R 2 adj. rrmse (%) D DART 0.908**** **** cm dbh cm dbh 0.773**** **** 14.3 DQ DART 0.930**** **** cm dbh 0.687*** *** cm dbh 0.388* ** 5.2 BA.10 cm dbh 0.741**** **** cm dbh 0.722*** **** 18.3 AGB.10 cm dbh 0.745**** **** cm dbh 0.723*** **** 18.8 CD DART 0.928**** **** 5.0 H DART 0.926**** **** 5.3 Notes: DART is the discrete anisotropic radiative transfer model. Parameters are: D, density (trees/ha); DQ, quadratic mean diameter (cm); BA, basal area (m 2 /ha); AGB, dry matter aboveground biomass (Mg/ha; 1 Mg ¼ 1 metric ton); CD, mean crown diameter (m); H, mean tree height (m). Adjusted (adj.) R 2 values and relative root mean square error (rrmse) are shown. Computed from the tree of mean height for simulated DART stands. * P 0.05; ** P 0.01; *** P 0.001; **** P predict forest biomass estimated from ground measurements using our regional allometric model (Fig. 2). Accuracy of GE results appeared visually to be slightly less stable than that of IKONOS results, and this was confirmed by a larger standard deviation of absolute residuals. GE images also gave a maximal relative error of 47% (i.e., 130 Mg for plot 3) against 31% (i.e., 41 Mg for plot 7) for IKONOS. Nonetheless, GE provides reliable predictions with a residual error very similar to the IKONOS-derived model (RMSE ¼ 80 and 77 Mg/ha, respectively), and no obvious saturation trend was observed above biomass values as high as 500 Mg/ha. Mapping forest aboveground biomass We used GE calibrated texture indices to estimate and map AGB values throughout the study area (Fig. 3). The position of the simulated DART scene of coarsest texture (i.e., with apparent crown diameter of 25 m) along the first PCA axis was taken as being a theoretical limit of the FOTO gradient validity domain. Beyond this limit, canopy grain is likely to result more from the presence of canopy gaps or topography effects, leading the model to predict abnormally high biomass values. The spatial trends of predicted AGB variability in Fig. 3 were consistent with variations in forest structure observed in the field. The general pattern showed an increase in biomass with elevation, corresponding to a decrease in accessibility from the Uppangala village (top-left corner) to the top of the major hills in the southern and eastern parts of the area. The lowest AGB estimates (i.e., darkest gray color), forming a diagonal band on the fully sunlit hills in the top part of the study area, are consistent with the occurrence of postfire, highly degraded secondary successions of very low biomass. DISCUSSION Consistent with Amazonian lowland evergreen terra firme (Couteron et al. 2005, Barbier et al. 2009) and mangrove forests (Proisy et al. 2007), the upper canopy of the wet evergreen forests of Indian WG shows pseudo-periodic patterns that allowed us to characterize and discriminate forest types on the basis of their canopy grain features. Our systematic analysis of canopy scenes by Fourier-based textural ordination revealed strong relationships between canopy texture and estimated stand structure parameters. The method does not suffer from saturation above biomass levels as high as 500 Mg/ha, and allowed us to build the first reliable map of tropical forest aboveground biomass predicted from free Google Earth images. FOTO potential to predict forest AGB The relationship between canopy texture and estimated stand structure has been the subject of little investigation due to the scarcity of relevant data sets, even though it could be a key step toward forest monitoring using very-high-resolution optical data. The FOTO indices, which describe the canopy texture through the breakdown of image variance with respect to Fourier periodic templates, were found in our study to strongly correlate with apparent crown diameter in the virtual canopy scenes simulated with DART (R ; see also Barbier et al. 2009, 2011a) as well as with canopy tree density (for trees with dbh. 30 cm, R ) from real sampled stands. Because the number of trees in the canopy is generally inversely related to their crown diameters, it is easily conceivable that the largest canopy trees play a key role in determining image texture properties. Yet texture provides structural information not only on the upper stand canopy but also on certain total stand characteristics. For instance,

7 April 2012 BIOMASS ASSESSMENT FROM CANOPY IMAGES 999 FIG. 2. Comparison between aboveground biomass (AGB, dry matter per ha; 1 Mg ¼ 1 metric ton) estimated from field plot measurements using a regional allometric model and from FOTO analysis of Google Earth (left) and IKONOS (right) canopy images of a wet evergreen forest in the Western Ghats of India. we demonstrated that FOTO indices determined for WG forests from GE and IKONOS canopy images also correlate with total stand basal area (R ) and AGB (R ). Also, in uneven-aged stands, although canopy trees only represent a small fraction of the stand population, they account for most of the total stand AGB. It therefore follows that although total density is difficult to estimate from a top-of-the-canopy view, total stand AGB may still be predictable, albeit at the cost of greater prediction error due to undetectable among-plot variations in the under-canopy contribution to AGB. However, the astonishing stability of our BA and AGB estimates when the focus was shifted from the subset of canopy trees to the entire stand population, FIG. 3. Left: Grayscale-converted Google Earth image of the study area in the Western Ghats of India. Right: dry matter aboveground biomass estimations (grayscale key with 95% confidence intervals for reference levels in parentheses) derived from FOTO canopy texture indices; study plot locations are the white-outlined squares. Some unit windows were excluded from analysis because they were masked prior to analysis (without grayscale code) or were filtered out because of topography-induced macroheterogeneous illumination (horizontally hatched squares), or were considered outside the model s validity domain due to very coarse canopy texture (diagonally hatched squares).

8 1000 PIERRE PLOTON ET AL. Ecological Applications Vol. 22, No. 3 points to the existence of general biophysical constraints on the spatial organization and development of forests stands. The top-down control of dominant trees capturing most of limiting resources (e.g., light) and framing energy/gas exchanges with the atmosphere over stand dynamics may allow the entire stand structure, and perhaps its functioning, to be determined from knowledge of merely a subsample of canopy trees (Enquist et al. 2009). An investigation of this key concept based on simulations of panchromatic images of more realistic 3-D forest mockups is underway (see Barbier et al. 2011a). Our correlation with total density of trees with dbh. 10 cm, however, contrasts with that obtained in the lowland evergreen forests of French Guiana (Couteron et al. 2005), which notably showed tree density to decrease significantly as canopy grain size increased. This may be due to the fact that our field data encompass an array of forest successions, from patches of highly degraded semievergreen secondary forests (resulting from an intense fire disturbance and yielding the finest canopy grain) to oldgrowth undisturbed evergreen stands with a coarse canopy grain. By contrast, the gradient of forest structure considered in the French Guiana case study concerned mature, undisturbed forests growing on contrasting soil conditions, and this resulted in marked differences in the sizes of the canopy trees (Couteron et al. 2005). Conversely, total basal area (BA) showed little variation across soil types, whereas significant variation was noted in terms of density, mean quadratic diameter, and canopy height. Density therefore does not always show predictable patterns with forest age or successional stages (e.g., Chazdon et al. 2007, Ramesh et al. 2010), unlike BA, which appears to be more representative of biomass accumulation over time (Chazdon et al. 2010). In both studies, however, mean quadratic diameter consistently showed high correlations with canopy texture indices, and thus appears to be a good candidate to bridge the gap between ground- and space-measured variables. Several studies have pointed to a strong allometric relationship between trunk section and crown surfaces in broad-leaved forests (Muller-Landau et al. 2006, Poorter et al. 2006), although more studies would be needed to gain a deeper understanding of this issue. Another asset of the FOTO method is its relative robustness with regard to spatial heterogeneity, at least once the most macro-heterogeneous images have been filtered out. At our study site, the most striking source of macro-heterogeneity is local topographic variations, which can create very large patterns of illumination that induce window-scale texture variations unrelated to variations in canopy properties. This local topographyinduced bias affected,5% of the total window set that was easily filtered out by a visual inspection prior to PCA. However, less obvious heterogeneity factors such as the juxtaposition of canopy gaps and clusters of aggregated crowns may yield window r-spectra that are skewed toward low frequencies, i.e., toward patterns larger than realistic maximum apparent crown diameter, therefore biasing the inference of stand structure parameters. Benchmarking and framing the texture gradient with DART scenes of known crown diameter distributions efficiently excluded the corresponding windows. We therefore provided realistic AGB predictions for 90% of the study area, which demonstrates the robustness and efficiency of the FOTO method. In this study, instrumental conditions and, in particular, sun and sensor geometries (Barbier et al. 2011b) were stable, and did not need to be accounted for. The potential large-scale textural variation introduced by slope aspect-related variations in sun-scene-sensor geometry was low. However, this latter effect, along with possible modifications in canopy structure on steep slopes, should be more thoroughly investigated in the future. Potential of free GE canopy images VHR Google Earth data are still largely unexploited by the scientific community (Potere 2008). Their potential for characterizing forest stand structure through texture analysis was compared here in this study with the analysis of native IKONOS images. We showed that GE images can be used for the consistent ordination of forest scenes on the basis of canopy textural properties, and that they are also suitable for estimating stand structure parameters and biomass with ranges of error comparable to those of IKONOS commercial images. The near infrared band (NIR), which has a great impact on the IKONOS panchromatic channel, is lacking in GE image spectral content, and this undoubtedly contributes to the observed texture differences between the two image types. Nonetheless, the fine level of spatial details given by GE images where VHR is available provides an interesting alternative to commercial images for texture-based methods, such as FOTO, that do not require preliminary radiometric corrections. The rapidly expanding spatial and temporal VHR coverage of GE images also offers a good potential for achieving broadscale results (see Barbier et al. 2009). The main shortcoming of GE data compared to commercial data is their higher geolocation error, estimated to range from 2 to 115 m in Western and South-Central Asia (Potere 2008). Although the positional error does not affect the spatial arrangement of image features (Montesano et al. 2009) and therefore does not bias the texture analysis, it hinders any accurate location of field plots in the image (McRoberts 2010) and thus model calibration. Method sensitivity to this common source of uncertainty in remote-sensing applications may be quantified by simple statistical measures (e.g., Asner et al. 2009) to better assess GE data limits. Overall, our results call for an in-depth study of the potential of GE images for FOTO, notably by (1) optimizing the image extraction procedure from GE in order to improve the spatial resolution that results from a trade-off between extent and altitude of the user view,

9 April 2012 BIOMASS ASSESSMENT FROM CANOPY IMAGES 1001 and (2) assessing the stability of textural gradients derived from GE with respect to a reference texture gradient provided by native full-resolution commercial VHR (e.g., IKONOS) images. This latter research avenue opens the perspective to run texture analyses on mixes of GE and commercial VHR images, an option that, in spite of the rapidly expanding spatial and temporal VHR coverage of GE images, would be very useful for conducting periodic affordable REDD-scale assessments of tropical forest degradation. Challenges facing accurate regional AGB estimations We attempted to quantify the main sources of error associated with the different steps in the strategy that we used to compute large-scale estimates of AGB for the wet evergreen forest of the WG. The texture AGB relationship yielded,15% relative error of the mean, which is fairly similar to that of lidar AGB relationships, generally reported to be ;20% (Goetz and Dubayah 2011). This confirmed the great potential of VHR optical data in characterizing the AGB of dense tropical forests, although not directly relevant to information on vertical stand structure. However, it is often overlooked that this remote-sensing-related error adds to other sources of error that propagate into the final AGB estimates (see Chave et al. 2004). Fieldmeasurement errors are uncorrelated with FOTO indices (i.e., PCA axes) and consequently cumulate with texture-related error. Although the design of our study precluded any quantification of field errors, three main sources of uncertainty should be considered. First, model selection error associated with the sample size of trees used to calibrate the allometry reflects our partial knowledge of tropical tree allometries. From an abacus obtained by Chave et al. (2004) using a rarefaction technique on a large data set, we evaluated this error to be ;5% of the mean AGB in our case, i.e., for ;200 trees as sampled in Rai s (1981) data set. A second source of error lies in the use of our allometric model beyond its range of validity (Chave et al. 2004). The maximum dbh in Rai s data set (i.e., 61 cm) is far smaller than that of the largest trees encountered in WG old-growth forests (up to 160 cm dbh). However, because only 4.5% of the sampled trees exceeded 61 cm dbh, we extrapolated large-tree AGB estimates from the same model. This, however, is an issue that should be specifically addressed in the future because very large trees can sometimes comprise the bulk of AGB. Lastly, a previous study conducted at UPSP demonstrated that the tree structure of some species was subjected to a topography effect, with individuals of a given diameter being shorter on steep than on gentle slopes (Robert and Moravie 2003). Such a difference in tree shape is likely to impact the ground-derived AGB estimates because our allometric model does not include tree height. At the tree level, the established allometric relationship linking tree diameter to its AGB was also subject to marked uncertainty due to a model error that we computed to be ;28%, to which should be added tree diameter measurement errors generally reported as ;5% (Chave et al. 2004). Although this source of error is often considered to cancel out at the plot level, it would be worth investigating ways to mitigate it with affordable increases in field efforts. For instance, it has been shown that including species- or even stand-level averages of wood specific density in AGB prediction models significantly improves their accuracy (Baker et al. 2004, Chave et al. 2005). No such data were available for the present study and the strategy of using a single wood density value across plots was rejected, given the suspected wide variations of mean local wood densities between plots located on disturbed softwood-dominated and old-growth hardwood-dominated stands. However, family- or genus-level averages have proved to significantly enhance allometric model performance (Baker et al. 2004) and the compilation of wood specific gravity databases dedicated to tropical tree species should therefore be encouraged. Tree height is also known to significantly improve the accuracy of allometric AGB models (Chave et al. 2004), but its individual ground measurement is dissuasively cumbersome and inaccurate for routine use. A solution might be to measure tree height in a subsample of trees (Couteron et al. 2005, Asner et al. 2010) so as to fit plot-specific diameter height allometries. On the other hand, lidar has recently emerged as an alternative remote-sensing technique for assessing forest biomass from canopy height information (Asner et al. 2010). Given the present unaffordable cost of systematically operating a lidar sensor to monitor forest degradation in the tropics, such data could be acquired over a limited number of selected sampling sites for calibrating accurate context-dependent reference AGB canopy texture relationships. ACKNOWLEDGMENTS We are grateful to the French Institute of Pondicherry (India) for providing logistic support for the project, and to the AMAP research unit (IRD, France) for funding. We also warmly thank the field team for its invaluable assistance in collecting data in the WG wet forests, namely, N. Ayyappan, Q. Renard, S. Ramalingam, T. Gopal, K. Adimoolam, and the villagers of Uppangala (India). This study falls within the framework of a joint research project between AMAP and IIRS (Indian Institute of Remote Sensing) supported by IFPCAR (Indo-French Promotion Center for Advanced Research) through grant 4409-C. LITERATURE CITED Asner, G. P., R. F. Hughes, T. A. Varga, D. E. Knapp, and T. Kennedy-Bowdoin Environmental and biotic controls over aboveground biomass throughout a tropical rain forest. Ecosystems 12: Asner, G. P., M. Palace, M. Keller, R. Pereira, Jr, J. N. M. Silva, and J. C. Zweede Estimating canopy structure in an Amazon forest from laser range finder and IKONOS satellite Observations. Biotropica 34: Asner, G. P., G. V. N. Powell, J. Mascaro, D. E. Knapp, J. K. Clark, J. Jacobson, T. Kennedy-Bowdoin, A. Balaji, G. Paez- Acosta, and E. Victoria High-resolution forest carbon stocks and emissions in the Amazon. Proceedings of the National Academy of Sciences 107:16738.

10 1002 PIERRE PLOTON ET AL. Ecological Applications Vol. 22, No. 3 Baker, T. R., et al Variation in wood density determines spatial patterns in Amazonian forest biomass. Global Change Biology 10: Barbier, N., P. Couteron, J.-P. Gastellu-Etchegorry, and C. Proisy. 2011a. Linking canopy images to forest structural parameters: potential of a modeling Framework. Annals of Forest Science. Barbier, N., P. Couteron, C. Proisy, Y. Malhi, and J.-P. Gastellu Etchegorry The variation of apparent crown size and canopy heterogeneity across lowland Amazonian forests. Global Ecology and Biogeography 19: Barbier, N., C. Proisy, C. Ve ga, D. Sabatier, and P. Couteron. 2011b. Bidirectional texture function of high resolution optical images of tropical forest: an approach using LiDAR hillshade simulations. Remote Sensing of Environment 115: Broadbent, E. N., G. P. Asner, M. Peña-Claros, M. Palace, and M. Soriano Spatial partitioning of biomass and diversity in a lowland Bolivian forest: Linking field and remote sensing measurements. Forest Ecology and Management 255: Brown, S., A. J. R. Gillepsie, and A. E. Lugo Biomass estimation methods for tropical forests with applications to forest inventory data. Forest Science 35: Chave, J., et al Tree allometry and improved estimation of carbon stocks and balance in tropical forests. Oecologia 145: Chave, J., R. Condit, S. Aguilar, A. Hernandez, S. Lao, and R. Perez Error propagation and scaling for tropical forest biomass estimates. Philosophical Transactions of the Royal Society B 359: Chazdon, R. L., B. Finegan, R. S. Capers, B. Salgado-Negret, F. Casanoves, V. Boukili, and N. Norden Composition and dynamics of functional groups of trees during tropical forest succession in northeastern Costa Rica. Biotropica 42: Chazdon, R. L., S. G. Letcher, M. van Breugel, M. Martínez- Ramos, F. Bongers, and B. Finegan Rates of change in tree communities of secondary Neotropical forests following major disturbances. Philosophical Transactions of the Royal Society 362: Couteron, P Quantifying change in patterned semi-arid vegetation by Fourier analysis of digitised aerial photographs. International Journal of Remote Sensing 23: Couteron, P., R. Pe lissier, E. Nicolini, and D. Paget Predicting tropical forest stand structure parameters from Fourier transform of very high-resolution remotely sensed canopy images. Journal of Applied Ecology 42: DeFries, R., F. Achard, S. Brown, M. Herold, D. Murdiyarso, B. Schlamadinger, and C. de Souza, Jr Earth observations for estimating greenhouse gas emissions from deforestation in developing countries. Environmental Science and Policy 10: Drake, J. B., R. O. Dubayah, R. G. Knox, D. B. Clark, and J.-B. Blair Sensitivity of large-footprint lidar to canopy structure and biomass in a neotropical rainforest. Remote Sensing of Environment 81: Enquist, B. J., G. B. West, and J. H. Brown Extensions and evaluations of a general quantitative theory of forest structure and dynamics. Proceedings of the National Academy of Sciences USA 106:7046. Foody, G. M Remote sensing of tropical forest environments: towards the monitoring of environmental resources for sustainable development. International Journal of Remote Sensing 24: Frazer, G. W., M. A. Wulder, and K. O. Niemann Simulation and quantification of the fine-scale spatial pattern and heterogeneity of forest canopy structure: A lacunaritybased method designed for analysis of continuous canopy heights. Forest Ecology and Management 214: Gastellu-Etchegorry, J. P., E. Martin, and F. Gascon DART: a 3D model for simulating satellite images and studying surface radiation budget. International Journal of Remote Sensing 25: Goetz, S., and R. Dubayah Advances in remote sensing technology and implications for measuring and monitoring forest carbon stocks. Carbon Management 2: Hansen, M. C., S. V. Stehman, P. V. Potapov, T. R. Loveland, J. R. G. Townshend, R. S. DeFries, K. W. Pittman, B. Arunarwati, F. Stolle, and M. K. Steininger Humid tropical forest clearing from 2000 to 2005 quantified by using multitemporal and multiresolution remotely sensed data. Proceedings of the National Academy of Sciences USA 105:9439. Houghton, R. A Tropical deforestation as a source of greenhouse gas emissions. Pages in P. Mutinho and S. Schwartzman, editors. Tropical deforestation and climate change. Amazon Institute for Environmental Research (IPAM), Bele m, Brazil. Huete, A., K. Didan, T. Miura, E. P. Rodriguez, X. Gao, and L. G. Ferreira Overview of the radiometric and biophysical performance of the MODIS vegetation indices. Remote Sensing of Environment 83: Imhoff, M. L Radar backscatter and biomass saturation: Ramifications for global biomass inventory. IEEE Transactions on Geoscience and Remote Sensing 33: IPCC Contribution of Working Groups I, II and III to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change. R. K. Pachauri and A. Reisinger, editors. IPCC, Geneva, Switzerland. Loffeier, M Sylviculture et sylvigéne` se en forêt sempervirente du Coorg (sud-ouest de l Inde). Travaux de la Section Scientifique et Technique 26. French Institute, Pondicherry, India. Maniatis, D., and D. Mollicone Options for sampling and stratification for national forest inventories to implement REDDþ under the UNFCCC. Carbon Balance and Management 5:1 14. MathWorks MatLab version 6.5. MathWorks, Natick, Massachusetts, USA. McRoberts, R. E The effects of rectification and Global Positioning System errors on satellite image-based estimates of forest area. Remote Sensing of Environment 114: Montesano, P. M., R. Nelson, G. Sun, H. Margolis, A. Kerber, and K. J. Ranson MODIS tree cover validation for the circumpolar taiga tundra transition zone. Remote Sensing of Environment 113: Mougin, E., C. Proisy, G. Marty, F. Fromard, H. Puig, J. L. Betoulle, and J.-P. Rudant Multifrequency and multipolarization radar backscattering from mangrove forests. IEEE Transactions on Geoscience and Remote Sensing 37: Muller-Landau, H. C., et al Comparing tropical forest tree size distributions with the predictions of metabolic ecology and equilibrium models. Ecology Letters 9: Pascal, J.-P Wet evergreen forests of the Western Ghats of India: ecology, structure, floristic composition and succession. Travaux de la Section Scientifique et Technique 20. French Institute, Pondicherry, India. Pascal, J., and R. Pe lissier Structure and floristic composition of a tropical evergreen forest in southwest India. Journal of Tropical Ecology 12: Pélissier, R., J.-P. Pascal, N. Ayyappan, B. R. Ramesh, S. Aravajy, and S. R. Ramalingam Tree demography in an undisturbed Dipterocarp permanent sample plot at Uppangala, Western Ghats of India. Ecology 92:1376. Pélissier, R., J.-P. Pascal, H. Houllier, and H. Laborde Impact of selective logging on the dynamics of a low elevation dense moist evergreen forest in the Western Ghats

11 April 2012 BIOMASS ASSESSMENT FROM CANOPY IMAGES 1003 (South India). Forest Ecology and Management 105: Poorter, L., L. Bongers, and F. Bongers Architecture of 54 moist-forest tree species: traits, trade-offs, and functional groups. Ecology 87: Potere, D Horizontal positional accuracy of Google Earth s high-resolution imagery archive. Sensors 8: Proisy, C., P. Couteron, and F. Fromard Predicting and mapping mangrove biomass from canopy grain analysis using Fourier-based textural ordination of IKONOS images. Remote Sensing of Environment 109: Rai, S. N Productivity of tropical rain forests of Karnataka. Dissertation. University of Bombay, Bombay, India. Rai, S., and J. Proctor Ecological studies on four rainforests in Karnataka, India. I. Environment, structure, floristics and biomass. Journal of Ecology 74: Ramesh, B. R., P. D. Venugopal, R. Pe lissier, S. V. Patil, M. H. Swaminath, and P. Couteron Mesoscale patterns in the floristic composition of forests in central Western Ghats, India. Biotropica 42: Read, J. M., D. B. Clark, E. M. Venticinque, and M. P. Moreira Application of merged 1-m and 4-m resolution satellite data to research and management in tropical forests. Journal of Applied Ecology 40: Ripley, B. D Spatial statistics. John Wiley, Hoboken, New Jersey, USA. Robert, A Simulation of the effect of topography and tree falls on stand dynamics and stand structure of tropical forests. Ecological Modelling 167: Robert, A., and M.-A. Moravie Topographic variation and stand structure heterogeneity in a wet evergreen forest of India. Journal of Tropical Ecology 19: St-Onge, B., Y. Hu, and C. Vega Mapping the height and above-ground biomass of a mixed forest using lidar and stereo Ikonos images. International Journal of Remote Sensing 29: Thenkabail, P. S., E. A. Enclona, M. S. Ashton, C. Legg, and J. De Dieu Hyperion, IKONOS, ALI, and ETMþ sensors in the study of African rainforests. Remote Sensing of Environment 90: UNFCCC Copenhagen Accord. Conference of the Parties, Fifteenth session, Copenhagen, 7 18 December. United Nations Framework Convention on Climate Change FCCC/2009/L.7. eng/l07.pdf Van der Werf, G. R., D. C. Morton, R. S. DeFries, J. G. J. Olivier, P. S. Kasibhatla, R. B. Jackson, G. J. Collatz, and J. T. Randerson CO 2 emissions from forest loss. Nature Geoscience 2: Wulder, M. A., E. F. Le Drew, S. E. Franklin, and M. Lavigne Aerial image texture information in the estimation of northern deciduous and mixed wood forest leaf area index (LAI). Remote Sensing of Environment 64: Zhou, J., C. Proisy, X. Descombes, I. Hedhli, N. Barbier, J. Zerubia, J.-Ph. Gastellu-Etchegorry, and P. Couteron Tree crown detection in high resolution optical and LiDAR images of tropical forest. Article 78240Q in C. M. U. Neale and A. Maltese, editors. Remote sensing for agriculture, ecosystems, and hydrology XII. Proceedings of SPIE, International Society for Optical Engineering. org/ / SUPPLEMENTAL MATERIAL Appendix A Values of control structural parameters (Ecological Archives A A1). Appendix B Figure of the FOTO results obtained from IKONOS canopy windows (Ecological Archives A A2). Supplement Rai s (1981) tree biomass database as used in the main paper (Ecological Archives A S1).

12

13

14

15

16

17 ID Species DBH H SW BBW SBW LTW AGB 1 Dimocarpus longan Garcinia morella Garcinia gummi guta Garcinia gummi guta Dimocarpus longan Reinwardtiodendron anamalayanum Dimocarpus longan Reinwardtiodendron anamalayanum Dimocarpus longan Garcinia gummi guta Reinwardtiodendron anamalayanum Garcinia morella Reinwardtiodendron anamalayanum Reinwardtiodendron anamalayanum Dimocarpus longan Carallia brachiata Reinwardtiodendron anamalayanum Reinwardtiodendron anamalayanum Garcinia morella Garcinia morella Calophylium polyanthum Garcinia morella Garcinia morella Garcinia morella Garcinia morella Garcinia morella Reinwardtiodendron anamalayanum Calophylium polyanthum Garcinia gummi guta Garcinia gummi guta Garcinia morella Dimocarpus longan Garcinia gummi guta Garcinia morella Dimocarpus longan Garcinia gummi guta Reinwardtiodendron anamalayanum Garcinia gummi guta Dimocarpus longan Garcinia morella Garcinia gummi guta Garcinia morella Garcinia morella Dimocarpus longan Garcinia gummi guta Calophylium polyanthum Garcinia morella Dimocarpus longan Dimocarpus longan Garcinia morella Garcinia morella Dimocarpus longan Garcinia gummi guta Reinwardtiodendron anamalayanum Canarium strictum Garcinia gummi guta Reinwardtiodendron anamalayanum Canarium strictum Syzygium gardneri Dimocarpus longan Syzygium gardneri Holigarna beddomei & grahamii Reinwardtiodendron anamalayanum Dimocarpus longan Reinwardtiodendron anamalayanum Garcinia gummi guta Garcinia gummi guta Reinwardtiodendron anamalayanum Canarium strictum Canarium strictum Carallia brachiata Calophylium polyanthum Dimocarpus longan Garcinia gummi guta Reinwardtiodendron anamalayanum Dimocarpus longan Dipterocarpus indicus Garcinia gummi guta Garcinia gummi guta Dipterocarpus indicus Dimocarpus longan

18 82 Reinwardtiodendron anamalayanum Dimocarpus longan Canarium strictum Dipterocarpus indicus Garcinia morella Calophylium polyanthum Persea macrantha Holigarna beddomei & grahamii Reinwardtiodendron anamalayanum Garcinia gummi guta Dipterocarpus indicus Reinwardtiodendron anamalayanum Canarium strictum Persea macrantha Holigarna beddomei & grahamii Carallia brachiata Garcinia gummi guta Persea macrantha Persea macrantha Carallia brachiata Dipterocarpus indicus Holigarna beddomei & grahamii Syzygium gardneri Persea macrantha Persea macrantha Persea macrantha Holigarna beddomei & grahamii Canarium strictum Garcinia morella Syzygium gardneri Syzygium gardneri Canarium strictum Persea macrantha Dipterocarpus indicus Dipterocarpus indicus Persea macrantha Dimocarpus longan Canarium strictum Holigarna beddomei & grahamii Carallia brachiata Dipterocarpus indicus Garcinia gummi guta Garcinia morella Dipterocarpus indicus Canarium strictum Canarium strictum Syzygium gardneri Canarium strictum Reinwardtiodendron anamalayanum Dimocarpus longan Persea macrantha Syzygium gardneri Holigarna beddomei & grahamii Dipterocarpus indicus Syzygium gardneri Calophylium polyanthum Carallia brachiata Calophylium polyanthum Persea macrantha Reinwardtiodendron anamalayanum Holigarna beddomei & grahamii Calophylium polyanthum Persea macrantha Carallia brachiata Calophylium polyanthum Carallia brachiata Persea macrantha Holigarna beddomei & grahamii Persea macrantha Syzygium gardneri Carallia brachiata Canarium strictum Calophylium polyanthum Holigarna beddomei & grahamii Calophylium polyanthum Persea macrantha Holigarna beddomei & grahamii Persea macrantha Holigarna beddomei & grahamii Persea macrantha Syzygium gardneri Calophylium polyanthum

19 164 Calophylium polyanthum Canarium strictum Holigarna beddomei & grahamii Holigarna beddomei & grahamii Canarium strictum Holigarna beddomei & grahamii Holigarna beddomei & grahamii Canarium strictum Calophylium polyanthum Canarium strictum Persea macrantha Persea macrantha Calophylium polyanthum Calophylium polyanthum Canarium strictum Holigarna beddomei & grahamii Holigarna beddomei & grahamii Carallia brachiata Holigarna beddomei & grahamii Calophylium polyanthum Persea macrantha Holigarna beddomei & grahamii Canarium strictum Calophylium polyanthum Calophylium polyanthum Calophylium polyanthum