1
Mapping the structure of Borneo’s tropical forests across a degradation gradient 1
2
Pfeifer M1*, Kor L1, Nilus R2, Turner E3, Cusack J4, Lysenko I1, Khoo M1, Chey VK2, Chung 3
AC2, Ewers RM1 4
5
1 Department of Life Sciences, Imperial College London, Silwood Park Campus, Buckhurst 6
Road, Ascot, Berkshire SL5 7PY, UK. [email protected]; 7
[email protected]; [email protected]; [email protected]; 8
2 Forest Research Centre, Sabah Forestry Department, PO Box 1407, 90715 Sandakan, 10
Sabah, Malaysia. [email protected] 11
3 Insect Ecology Group, Department of Zoology, University of Cambridge, Downing Street, 12
Cambridge, CB2 3EJ, UK. [email protected] 13
4 Department of Zoology, University of Oxford, The Tinbergen Building, South Parks Road, 14
Oxford, OX1 3PS, UK. [email protected] 15
16
* Corresponding author 17
18
Keywords: Borneo, humid tropical forests, degradation, RapidEyeTM, co-occurrence 19
measures, spectral reflectance, biophysical structure, maps, field data20
2
Abstract 21
South East Asia has the highest rate of lowland forest loss of any tropical region, with 22
logging and deforestation for conversion to plantation agriculture being flagged as the most 23
urgent threats. Detecting and mapping logging impacts on forest structure is a primary 24
conservation concern, as these impacts feed through to changes in biodiversity and ecosystem 25
functions. Here, we test whether high-spatial resolution satellite remote sensing can be used 26
to map the responses of aboveground live tree biomass (AGB), canopy leaf area index (LAI) 27
and fractional vegetation cover (FCover) to selective logging and deforestation in Malaysian 28
Borneo. We measured these attributes in permanent vegetation plots in rainforest and oil 29
palm plantations across the degradation landscape of the Stability of Altered Forest 30
Ecosystems project. We found significant mathematical relationships between field-measured 31
structure and satellite-derived spectral and texture information, explaining up to 62 % of 32
variation in biophysical structure across forest and oil palm plots. These relationships held at 33
different aggregation levels from plots to forest disturbance types and oil palms allowing us 34
to map aboveground biomass and canopy structure across the degradation landscape. The 35
maps reveal considerable spatial variation in the impacts of previous logging, a pattern that 36
was less clear when considering field data alone. Up-scaled maps revealed a pronounced 37
decline in aboveground live tree biomass with increasing disturbance, impacts which are also 38
clearly visible in the field data even a decade after logging. Field data demonstrate a rapid 39
recovery in forest canopy structure with the canopy recovering to pre-disturbance levels a 40
decade after logging. Yet, up-scaled maps show that both LAI and FCover are still reduced in 41
logged compared to primary forest stands and markedly lower in oil palm stands. While 42
uncertainties remain, these maps can now be utilised to identify conservation win-wins, 43
especially when combining them with ongoing biodiversity surveys and measurements of 44
carbon sequestration, hydrological cycles and microclimate. 45
Deleted: ¶46
3
1. Introduction 47
Through the processes of selective logging and agricultural conversion, large areas of tropical 48
forests are being degraded and fragmented worldwide (Blaser, Sarre, Poore, & Johnson, 49
2011). South East Asia, in particular, has the highest rate of lowland forest loss of any 50
tropical region, with deforestation and logging for conversion to oil palm being flagged as the 51
most urgent threats (Sodhi, Koh, Brook, & Ng, 2004). Conventional logging, widely used in 52
South East Asia, and to a lesser extent ‘Reduced Impact Logging’ (Pinard & Putz, 1996) 53
reduces aboveground biomass (AGB) through the removal of large trees (Slik et al., 2013), 54
and causes residual damage that includes opening previously dense forest canopies (Pfeifer et 55
al., 2015). Resultant decreases in canopy leaf area index (LAI) and fractional vegetation 56
cover (FCover) affect microclimate (Didham & Lawton, 1999; Hardwick et al., 2015), carbon 57
sequestration through photosynthesis and regulation of surface water flows (Douglas, 1999). 58
These changes can affect biodiversity (Edwards et al., 2011; Wilcove, Giam, Edwards, 59
Fisher, & Koh, 2013) and ecosystem functions operating at different trophic levels (Ewers et 60
al., 2015). 61
Thus, unsurprisingly, detecting and mapping forest degradation is a primary conservation 62
concern, and remote sensing could provide an effective means to achieve this task (Rose et 63
al., 2014). Satellite or airborne sensor data have previously been employed to map 64
deforestation at coarse to fine spatial resolutions from landscape to global scales (Hansen et 65
al., 2013). Sensor-data based forest degradation and change maps combined with ground-66
surveys aid in quantifying degradation impacts on carbon stocks and emissions (Baccini et 67
al., 2012; Harris et al., 2012), as well as on biodiversity (Gibson et al., 2011; Strassburg et al., 68
2010; Wearn, Reuman, & Ewers, 2012). 69
Yet, mapping forest degradation using space-borne sensors is challenging in many 70
tropical humid landscapes, not least because of persistent cloud cover. Tree cover maps 71
4
(Hansen et al. 2013) or simpler maps of vegetation greenness (Pettorelli et al., 2005) do not 72
directly map forest biophysical structure or biomass and struggle to distinguish primary from 73
disturbed forests or agricultural land, including oil palm stands (Tropek et al., 2014). The 74
latter now cover large areas of South East Asian landscapes (Koh & Wilcove, 2008), and are 75
easily confused with forests during image classifications, especially when reaching maturity. 76
A number of methods can distinguish logged from primary forests (Asner et al., 2005). 77
But, these methods map categories and thus treat the habitat within each land-use as 78
homogenous. Yet, forest canopy structure and biomass can vary considerably within logged 79
and unlogged forests affecting ecosystem processes on a continuous scale. For instance, the 80
intensity of forest disturbance is a key determinant of the carbon (Berenguer et al., 2014; 81
Pfeifer et al., 2015) and biodiversity (Burivalova, Şekercioǧlu, & Koh, 2014; Gibson et al., 82
2011) that persists in logged forest. Carbon and biodiversity, in turn, are used to dictate 83
conservation and management priorities at the landscape scale under certification schemes 84
such as the High Carbon Stock and High Conservation Value standards applied to the palm 85
oil industry (Greenpeace International, 2013; RSPO, 2013). 86
Successfully linking sensor data to actual forest structure data is challenging. Multi-87
sensor approaches combining airborne LIDAR with satellite data are delivering promising 88
results in estimating canopy gap structure at landscape scales (Asner, 2009; Asner et al., 89
2010). But despite the decreasing costs of airborne LIDAR data (Asner, 2009), their 90
widespread use is still limited by the need for expensive expert knowledge for data 91
processing, large financial resources to fund planes and flights, and special flying permits that 92
are often hard to obtain in many tropical countries. Spaceborne Synthetic Aperture Radar 93
(SAR) such as ALOS PALSAR overcomes the challenges of persistent cloud cover and 94
atmospheric effects when mapping forest degradation (Woodhouse, Mitchard, Brolly, 95
Maniatis, & Ryan, 2012), with long-wavelength L-Band signals penetrating vegetation 96
5
canopies and relating to forest biophysical properties (Boyd & Danson, 2005; Lucas et al, 97
2014). However, costs of Radar imagery are high and availability limited (Joshi et al., 2015). 98
Backscatter starts to saturate at around 100 - 150 Mg ha-1 with errors being introduced largely 99
in ≥ 150 Mg ha-1 forest stands (Robinson, Saatchi, Neumann, & Gillespie, 2013; Mermoz et 100
al., 2015) typical for humid tropical forests with dense canopies. Also, backscattered energy 101
is not only affected by the size and orientation of the structural elements in the vegetation 102
canopy, but also by the moisture content of the vegetation and the underlying soil conditions 103
(Naidoo et al., 2015). 104
Unlike LIDAR and Radar, moderate to high-resolution satellite images can be obtained 105
for most regions in the world free of charge (e.g. Landsat 8 data via USGS Earth Explorer; 106
SPOT and RapidEyeTM data via European Space Agency). Spectral and texture information 107
derived from such images have been exploited by ecologists to map tree cover (Céline et al., 108
2013), green foliage density (Glenn, Huete, Nagler, & Nelson, 2008), and forest biomass and 109
canopy structure (Castillo-Santiago, Ricker, & de Jong, 2010; Pfeifer, Gonsamo, Disney, 110
Pellikka, & Marchant, 2012; Wang, Qi, & Cochrane, 2005). In addition, tools to process 111
these images are now easily accessible via statistical and spatial analyses software, including 112
freeware such as R Open Source Statistical software (R Development Core Team, 2013) and 113
QGIS Open Source Geographic Information System (QGIS Development Team, 2012), more 114
accessible to management staff in the tropics and elsewhere. 115
Here, we test whether significant relationships can be found between high spatial 116
resolution optical satellite sensor data and three attributes of forest structure (AGB, LAI, 117
FCover) measured in rainforest and oil palm plantation areas across a degradation landscape 118
in Malaysian Borneo. Satellite imagery was obtained at no-cost from the European Space 119
Agency under Category 1 User Agreement. Forest attributes were collected in permanent 120
vegetation plots (a combined survey area of > 12 ha) established across the Stability of 121
6
Altered Forest Ecosystems (SAFE) Project landscape (Ewers et al., 2011). In particular, we 122
ask: (1) what are the impacts of logging on forest structure more than one decade after 123
logging, and are these impacts spatially homogenous or heterogeneous; and (2) can we 124
reliably detect spatial variability in logging impacts, observed on the ground, using high 125
spatial resolution passive sensor data? Our aim was to identify reliable mathematical 126
relationships that would enable mapping of continuous surfaces of forest and oil palm 127
structure across entire landscapes. This would provide important baseline data for use in the 128
planning and management of future logging and agricultural activities in Borneo’s highly 129
threatened forests to minimise their environmental footprints. 130
131
2. Methods 132
133
2.1 The SAFE degradation landscape 134
The Stability of Altered Forest Ecosystem (SAFE) Project (4° 38’ N to 4° 46’ N, 116° 57’ to 135
117° 42’ E; Fig. 1), located within lowland dipterocarp forest regions of East Sabah in 136
Malaysian Borneo, features a gradient of forest disturbance from unlogged primary forest 137
through to severely degraded forest and oil palm plantations (Ewers et al., 2011). SAFE thus 138
reflects Sabah’s predominant land use change over the past decades, characterised by 139
industrial harvesting and premature re-logging of extensive tracts of once-logged forests and 140
large-scale forest-to-palm conversions (Reynolds, Payne, Sinun, Mosigil, & Walsh, 2011). 141
The region has a tropical climate with high rainfall (> 2000 mm / year) and varying terrain 142
topography, although all plots are below 800 m altitude. The geology comprises a mixture of 143
sedimentary rocks including siltstones, sandstones and others that are easily eroded (Douglas 144
et al. 1999). 145
7
Meteorological records from nearby Danum Valley Field Centre, show that the climate in 146
this part of Sabah is aseasonal, with occasional dry spells that are usually associated with El 147
Niño events (Walsh & Newbery, 1999). Recent data from Danum Valley (SEARRP: 148
http://www.searrp.org/danum-valley/the-conservation-area/climate) show that no dry months 149
occurred between September 2011 and January 2013, the period during which our data was 150
collected. This is important, as this could have potentially changed biophysical attributes 151
between time of field measurements and time of satellite data acquisition, as reported for 152
drought-affected areas in evergreen Amazonian forests (Myneni et al., 2007). 153
193 vegetation plots (25 m x 25 m) with North–South orientation were set out at SAFE in 154
2010 using a Garmin GPSMap60 device. Vegetation plots were established across the forest 155
degradation landscape according to a hierarchical sampling design as an objective procedure 156
to assess regional forest attributes (Ewers et al 2011). The design was chosen to ensure 157
unbiased decisions as to where to establish vegetation monitoring plots in the field (Ewers et 158
al 2011). Plots were located at roughly equal altitude and oriented to minimize potentially 159
confounding factors such as slope, latitude, longitude and distance to forest edges (prior to 160
controlled forest-to-oil palm conversion currently being carried out at SAFE) (Ewers et al 161
2011). Plots were distributed among 17 sampling blocks: three oil palm plantation blocks of 162
two different ages (OP1 and OP2 planted in 2006 and OP3 planted in 2000), two primary 163
forest stands (OG1 and OG2), lightly or illegally logged forests inside multi-use or protected 164
areas (OG3 and VJR), twice logged forests (LFE, LF1 - LF3) and salvage logged forest (A - 165
F). The latter is forest that has been logged with lifted restrictions on type of tree, size limits 166
and volume in advance of outright conversion to a new land use type such as agriculture (for 167
details on the hierarchical sampling design see Ewers et al., 2011). Twice and salvage logged 168
forests were selectively logged, once during the 1970s followed by a second logging round 169
from the late 1990s to the early 2000s removing medium hardwoods (Drybalanops and 170
Deleted: ¶171
Formatted: Font: 12 pt, Font color: Auto
http://www.searrp.org/danum-valley/the-conservation-area/climate
8
Dipterocarpus), and lighter hardwoods (Shorea and Parashorea). Historical forestry data 172
suggest that logging implemented under a modified uniform system removed an estimated 173
113 m3 ha-1 during the first rotation followed by an estimated 37 m3 ha-1 extraction during the 174
second rotation in lightly logged forests; the salvage logged forest were re-logged three times 175
with a cumulative extraction rate of 66 m3 ha-1 (Struebig et al., 2013), although our up-scaled 176
maps indicate even higher harvesting intensities. Overall, logging in the area resulted in 177
heavily degraded stands with a high density of roads and skid trails, a paucity of commercial 178
timber species, few emergent trees and the dominance of pioneer and invasive vegetation. At 179
SAFE, blocks A - F are currently being converted into a fragmented agricultural landscape. 180
181
2.2 Quantifying forest structure: AGB, LAI, FCover 182
Permanent vegetation structure plots were measured for above-ground biomass in 2010 and 183
2011 (N = 193). We measured canopy structure in 172 of these plots in 2012 and 2013. 184
Above-ground biomass was derived from measurements of tree size (height and diameter at 185
breast height, DBH) following RAINFOR protocols and including all trees with DBH ≥ 10 186
cm and with > 50 % of their roots located within the plot boundaries (for details see Pfeifer et 187
al., 2015). 188
We computed above-ground biomass using an improved pan-tropical algorithm (Chave et 189
al., 2014), as a recent study showed that it provided better biomass estimates than regional 190
algorithms for East Kalimantan (Rutishauser et al., 2013). We computed AGB using wood 191
specific gravity values reflecting different disturbance level of forest stands in Borneo (Slik et 192
al., 2008). For each tree, we computed AGB as above but using wood specific gravity 193
estimates drawn randomly (N = 100) from a normal distribution of wood specific gravity 194
values (for details see Pfeifer et al., 2015). This distribution was defined by maximum and 195
standard deviation and depended on the tree’s location: primary forests (0.64 ± 0.18) (King et 196
9
al 2006), and lightly logged forests (0.57 ± 0.02), twice logged forests (0.54 ± 0.03), and 197
salvage logged forests (0.41 ± 0.05) (Slik et al 2008). Oil palms have a fundamentally 198
different physical structure to forest trees, so we estimated AGB in oil palm plantations 199
separately (Pfeifer et al., 2015). We computed dry mass in Mg for each palm tree from its 200
height (in m) using sing (Thenkabail et al 2004). 201
Leaf area index (corrected for foliage clumping) and fractional vegetation cover were 202
estimated from 12 to 13 hemispherical photographs acquired in each plot following standard 203
sampling procedures (Pfeifer et al., 2014), using in house-algorithms (Pfeifer et al., 2014) and 204
the freeware CAN-EYE v 6.3.8 (Weiss & Baret, 2010). 205
206
2.3 Satellite data and derived spectral and texture data 207
We used five RapidEyeTM satellite images acquired over the SAFE landscape in 2012 and 208
January 2013 (Table A1 in Appendix). These cover the entire globe with a temporal 209
resolution of 5.5 days (at nadir) and are high-resolution, ortho-rectified products resampled to 210
5 m pixel resolution. The Multi-Spectral Imager on-board the satellites acquires data in five 211
spectral bands. The blue (0.44 - 0.51 µm), green (0.52 - 0.59 µm), red (0.63 - 0.68 µm) and 212
near-infrared (0.76 - 0.85 µm) bands are very similar to the corresponding Landsat spectral 213
bands, but the sensor has an additional red-edge band (0.69 - 0.73 µm), giving additional 214
sensitivity to changes in the reflectance of vegetation (Chander et al., 2013). 215
We pre-processed all images (calibration, cloud and cloud shadowing-masking) using 216
ENVI v 5.1 (Exelis Visual Information Solutions, Boulder, Colorado). We applied 217
atmospheric correction to these images using the 6s algorithm (Second Simulation of Satellite 218
Signal in the Solar Spectrum) (Vermote, Tanre, Deuze, Herman, & Morcette, 1997) 219
implemented with the Open Source Geographic Resources Analysis Support System 220
(GRASS) i.atcorr function (Neteler, Bowman, Landa, & Metz, 2012) implemented in 221
10
Quantum GIS v 2.6.1 (Hugentobler, 2008). We computed standard spectral vegetation 222
indices, including Normal Difference Vegetation Index (NDVI) (Tucker, 1979), a two-band 223
Enhanced Vegetation Index (EVI2) (Jiang, Huete, Didan, & Miura, 2008), and the Modified 224
Soil-Adjusted Vegetation Index (MSAVI2) (Qi, Chehbouni, Huete, Kerr, & Sorooshian, 225
1994), as well as texture indices using the ‘raster’ (Hijmans & van Etten, 2010) and ‘glcm’ 226
(Zvoleff, 2014) libraries in R statistical software (R Development Core Team, 2013). To 227
obtain texture indices, we computed a grey-level co-occurrence matrix (GLCM) with a 90 228
degree shift and 64 grey-levels and varied the window size for analyses from 3 x 3 to 9 x 9 229
pixels. This is important as the size of the analysis window used affects textural features, with 230
the most appropriate option depending on image resolution and landscape characteristics (Lu 231
2005). We computed the mean ( ), variance 232
( ), contrast ( ) and 233
dissimilarity ( ) as measures of image texture for each of the 234
five bands, where P is the symmetrical GLCM with dimension N x N and i and j are the ith 235
and jth element of P (Gallardo-Cruz et al. 2012). Pij therefore represents the probability of 236
finding the value of the reference pixel i in combination with the neighbouring pixel j. 237
We extracted spectral intensity and texture data for each band as well as spectral 238
vegetation indices to the point coordinates of each plot as mean values within a buffer radius 239
of 20 m. During this step, we excluded 58 plots that were covered by cloud and cloud 240
shadows. We used these values as predictors in subsequent modelling and removed highly 241
inter-correlated indices from global models. In line with other studies (Wang, Qi, & 242
Cochrane, 2005), MSAVI2 outperformed NDVI and EVI2 (Pearson's product moment 243
correlation computed for measures of vegetation greenness: r ≥ 0.9; see also Figure A1 in 244
Appendix) in its ability to separate primary from lightly-logged forest stands and was 245
subsequently used in all analyses. Dissimilarity estimates computed from band intensities 246
11
within a window of 9 x 9 pixels outperformed other texture indices in single predictor 247
regression models (see Tables A2 and A3 in Appendix), and we subsequently used this 248
resolution in all analyses. We excluded intensity and dissimilarity values for the blue band in 249
modelling and upscaling forest structure. Visual inspections revealed obvious striping and 250
banding in the blue band of some images that almost certainly derive from image calibration, 251
which is known to be problematic in large amount of cloud-filled transition scenes (Godard, 252
Stoll, Anderson, Schulze, & D’Souza, 2012). 253
254
2.4 Data analysis 255
Plot data (forest attributes and spectral as well as texture indices) were aggregated to the level 256
of the 17 sampling blocks and five disturbance levels defined as oil palm (OP1-OP3, three 257
blocks), salvage logged (A-F, six blocks), twice-logged (LFE & LF1-LF3, four blocks), 258
lightly logged (OG3 & VJR, two blocks) and primary forest (OG1& OG2, two blocks). We 259
used one-way ANOVA with post-hoc Tukey HSD tests for pairwise comparisons aimed at 260
detecting general differences among the three metrics of forest structure and among earth 261
observation signals with respect to disturbance level. We checked for homogeneity of 262
variance within groups using Bartlett's K-squared test statistic, and computed the Welsh 263
correction to compare pairwise group means in case of variance non-homogeneity using the 264
R ‘stats’ package (R Development Core Team, 2013). 265
We explored single predictor relationships between earth observation data and forest 266
structure data across the three levels of aggregation (plots, blocks, disturbance levels), using 267
nonlinear least squares and general linear models as implemented in the R ‘stats’ package (R 268
Development Core Team, 2013). 269
We used beta-logistic regression with the betareg package in R (Grün, Kosmidis, & 270
Zeileis, 2012; Simas, Barreto-Souza, & Rocha, 2010) to develop multiple predictor models. 271
12
We scaled model predictions between lower and upper bounds, which were set to reflect 272
assumptions about expected forest structure per plot (each 0.0625 ha); FCover: 0 - 100, LAI: 273
0 - 10, AGB: 0 - 65 (maximum value observed at site was 60.8 Mg / Plot). We fitted beta-274
logistic regression models for plot level data using Maximum Likelihood and automated 275
model selection via information theoretic approaches and multi-model averaging. First, we 276
constructed a global model each for LAI, FCover and AGB, including predictors from the set 277
of predictors described above, after first excluding highly correlated predictors (see Table 1 278
for details on global models). For each global model, we used the dredge function in the R 279
MuMIn package v1.10.5 (Barton 2014), which constructs models using all possible 280
combinations of the explanatory variables supplied in each global model. These models were 281
ranked, relative to the best model, based on the change in the Akaike Information Criterion 282
(delta AIC). A multi-model average (final model) was calculated across all models with delta 283
AIC < 4. 284
285
2.5 Upscaling structure across the landscape 286
We added the grey-level dissimilarity layers as well as the MSAVI2 layer to each 287
RapidEyeTM scene and subsequently mosaicked those rasters. We aggregated all rasters to 25 288
m pixel resolution maps, to match the scale of vegetation structure plots using the R raster 289
package (Hijmans & van Etten, 2010). We then used the model-averaged final model to 290
predict scaled mean and fitted quantiles of the response distribution with lower (at 0.25) and 291
upper quantiles (at 0.75) estimates of AGB, FCover and LAI for each pixel in the SAFE 292
landscape. We unscaled model estimates of forest structure traits and saved the output as new 293
rasters. 294
We created minimum convex polygons with a 200 m buffer enclosing all plots within 295
each block. We extracted mean AGB, LAI and FCover for each pixel encompassed by each 296
13
polygon and summarised overall patterns for variation in the forest structure attributes across 297
the five disturbance levels using density curves and histogram statistics. 298
299
3. Results 300
301
3.1 Variability of forest structure across the forest disturbance gradient 302
Oil palms differed structurally from primary and degraded forest, featuring significantly 303
lower biomass and canopy cover (Fig. 2). Structural differences were less consistent between 304
primary and disturbed forests. Above-ground biomass decreased significantly from primary 305
(Mean: 413.4 t / ha) to lightly logged (267.3 t / ha) to twice-logged (110.6 t / ha) and salvage 306
logged forest (42.4 t / ha). LAI and FCover were significantly higher in twice-logged versus 307
salvage logged forests (P < 0.05, Welsh corrected). We found no further significant 308
differences between forest types with regard to forest canopy structure (Fig. 2). 309
310
3.2 Spectral and texture attributes across the forest disturbance gradient 311
Oil palm blocks differed significantly from forests and displayed lower vegetation greenness 312
(except oil palm versus twice-logged forest), higher red intensity and dissimilarity and lower 313
near-infrared dissimilarity (except oil palm versus salvage logged forest) (Fig. 3). The red 314
intensity was lower in primary and lightly logged forest compared to salvage logged and 315
twice-logged forest, indicating higher photosynthetic capacity in less-disturbed stands (P < 316
0.05, Welsh corrected). Patterns were similar for the red-edge intensity but less pronounced 317
and not significant for oil palm versus salvage logged, and primary and lightly logged versus 318
twice logged forest. The near-infrared intensity was significantly higher in salvage logged 319
forest and oil palm compared to lightly and twice-logged forest (all P < 0.01, except oil palm 320
versus lightly logged: P = 0.07). Vegetation greenness was higher in primary compared to 321
14
twice-logged and salvage logged forest (P < 0.001, Welsh corrected) and lower in twice-322
logged versus salvage logged forest (P < 0.005). The red dissimilarity was higher in primary 323
and twice-logged forest compared to salvage logged forest (P < 0.001, Welsh corrected). In 324
addition, primary forest had significantly lower near-infrared band dissimilarity compared to 325
twice-logged forest (P < 0.05) but significantly higher near-infrared band dissimilarity 326
compared to salvage logged forest (Welsh corrected, P < 0.001). The near-infrared 327
dissimilarity was lower in salvage logged forest compared to twice-logged (P < 0.0001) and 328
lightly logged forest (Welsh corrected, P < 0.001). 329
330
3.3 Modelling relationships between forest attributes and remotely sensed measures 331
Clear relationships emerged between earth observation data and forest structure attributes 332
across aggregation levels (Fig. 4). At stand level, leaf area index and fractional vegetation 333
cover showed strong relationships with red intensity in single-predictor linear models (LAI = 334
8.556 - 1.059 * RED, RSS = 0.59, Deviance = 0.65, coefficient and intercept significant at P 335
< 0.0001; FCover = 152.09 - 17.59 * RED, RSS = 9.475, Deviance = 0.66, coefficient and 336
intercept significant at P < 0.0001). Aboveground biomass increased exponentially with 337
increasing vegetation greenness (AGB = 6.729e-17 * exp(51.18*MSAVI2), RSS = 4.845, 338
Deviance = 0.79, coefficient significant at P < 0.001). At plot level, relationships are weaker 339
but still significant (FCover = 124.918 -11.685 * RED, RSS = 16.13, Deviance = 0.24, 340
coefficient and intercept significant at P < 0.0001; LAI = 7.18 - 0.76 * RED, RSS = 0.8466, 341
Deviance = 0.33, coefficient and intercept significant at P < 0.0001; AGB = 4.701e-8 * 342
exp(24.52 *MSAVI2), RSS = 9.009, Deviance = 0.13, coefficient significant at P < 0.0001). 343
Including additional predictors improved model performance at the plot level (Table 2). 344
The green and red-edge intensities as well as vegetation greenness were most important 345
predictors of AGB (Table 2). Grey-level dissimilarities of red and red-edge (in the case of 346 Deleted: and near-infrared 347
15
LAI) and of red edge (for FCover) improved performance of models predicting forest canopy 348
attributes. Observed values showed reasonable agreement with predicted values for all three 349
forest attributes at the level of forest types, although variability at plot level is high (Figure 350
A2 in Appendix). Our models underestimate AGB in primary and lightly logged forests, and 351
overestimate canopy structure attributes in oil palms (Figure A2). It is important to note, 352
however, that the models describing the relationships between forest canopy structure and 353
sensor data were strongly influenced by the distinction between oil palm and forest habitats, 354
but that variation among forest plots was less strongly differentiated. 355
356
3.4 Up-scaled estimates of logging damage 357
Maps of forest structure attributes show that all forest types differ significantly from each 358
other in their structure (Fig. 5). On average, AGB (t / ha), LAI and FCover decreased 359
sequentially from primary forest to lightly logged, twice-logged and salvage logged forests 360
(Table 3). Field data as well as map-derived AGB estimates clearly showed the decline in 361
forest biomass with increasing disturbance (see also Pfeifer et al., 2015). Our maps indicate 362
average biomass extraction levels of 187 t / ha for lightly logged forest, 228 t / ha for twice-363
logged forest, and 255 t / ha for salvage logged forest. Yet, whilst field data suggest rapid 364
canopy recovery following disturbance, data extracted for buffered minimum convex 365
polygons from up-scaled maps reveal a more nuanced process. LAI was still reduced by 15 % 366
for salvage and by 7 % for lightly logged forest compared to primary forest. And forest types 367
differed significantly from each other (P < 0.001); FCover was still reduced by 12.4 % for 368
salvage logged forest and by 7.8 % for lightly logged forest and differed significantly 369
between forest types except between twice and lightly logged forests. 370
Biomass varied less in oil palm plantations compared to forests (Variance test: P < 0.001) 371
and varied more in primary forests compared to disturbed forest stands (P < 0.001); AGB 372
Deleted: FCover373
Deleted: LAI374
16
heterogeneity was lower in twice logged forests compared to little and salvage logged stands 375
(P < 0.001). Conversely, canopy LAI varied significantly more in oil palm plantations 376
compared to forests and in salvage logged forests compared to twice and lightly logged 377
forests (P < 0.001). However, canopy LAI varied significantly less in salvage logged stands 378
compared to primary forest stands. FCover varied significantly more in oil palms compared 379
to forests, in salvage logged compared to twice logged forests, in twice logged compared to 380
lightly logged forests; FCover varied more in primary compared to lightly logged forests but 381
less compared to other forest types (P < 0.001). 382
383
4. Discussion 384
Forest degradation via processes like selective logging, is a process that leads to a 385
deterioration in the density or structure of forest cover (Lambin, 1999) causing significant 386
loss of aboveground biomass (Slik et al., 2013) and reduced photosynthesis through the 387
removal of large trees and canopies (Figueira et al., 2008). Degradation impacts are 388
accompanied by structural changes at stand level, producing a heterogeneous forest landscape 389
comprised of intact forest, tree-fall gaps, remains of logging infrastructure with compacted 390
soils, and collateral damage on non-target trees (Laporte et al., 2007). Structural changes are 391
temporary, as productivity of smaller trees and saplings increases with light availability 392
allowing them to refill canopy gaps. Recovery time to pre-disturbance conditions of 393
biophysical structure, however, will vary with location and disturbance intensity (Frolking et 394
al., 2009; Martin et al., 2015). 395
Using our extensive network of field measurements, we show that canopy cover appears 396
to recover more rapidly than aboveground biomass in Borneo’s degraded forests. Canopy 397
structure and biomass are considerably lower in oil palm stands, though, indicating their 398
reduced functions in terms of carbon sequestration (Pfeifer et al., 2015) and microclimate 399
17
regulation (Hardwick et al., 2015). We found significant mathematical relationships between 400
ground measures of forest structure and high-spatial resolution satellite sensor data. These 401
relationships held at different aggregation levels, highlighting the potential to use these 402
relationships to map aboveground biomass, vegetation cover and canopy leaf area across the 403
SAFE degradation landscape (Fig. 6 and Figure A3 in Appendix). Such maps revealed 404
considerable spatial variability in the impacts of previous logging and forest-to-palm 405
conversion across the landscape. This pattern was less clear when considering field data 406
alone, despite the large survey area covered. 407
408
4.1 Spectral signatures across the degradation gradient 409
Higher absorption in the red wavelengths by canopies of undisturbed and lightly logged 410
forests indicates higher photosynthetic capacity compared to photosynthesis in forests that 411
were heavily logged more than ten years ago (see Fig. 3). Red absorption appears more 412
heterogeneous in primary forest versus lightly logged ones at stand level, possibly because 413
natural mortality of large trees that still exist in the undisturbed creates large canopy gaps, 414
visible in the high variability of canopy LAI. These gaps with intense regeneration are 415
surrounded by intact forests. This interpretation is supported by heterogeneous but high near-416
infrared reflectance measured above undisturbed forests. High near-infrared reflectance may 417
be caused by grass growth or strong regrowth of pioneer tree species, which are both patterns 418
we would expect in heavily disturbed stands or in recent larger canopy gaps (Riou & Seyler, 419
1997). Salvage logged forests are strongly degraded comprising regrowth patches covered in 420
less productive vines intersected by an extensive road network; they feature slow tree 421
regeneration and significantly reduced aboveground biomass. Thus, unsurprisingly, salvage 422
logged forests exhibit high but homogeneous near-infrared reflectance and low but 423
homogeneous red absorption. 424
Deleted: 2425
18
Twice logged stands differed clearly from undisturbed forest and dense forest regrowth 426
but also from salvage logged stands. They exhibit high near-infrared reflectance and low red 427
absorption, but both with variable texture patterns, which is indicative of rainforests with 428
signs of man-made disturbance (Riou & Seyler, 1997). These stands are irregular mosaics of 429
intact forest, logging infrastructure and clear-cuts that are currently regenerating. Oil palm 430
shows less red wavelength absorption and considerably higher dissimilarity in doing so, 431
probably because they vary in age and canopy LAI and hence photosynthetic capacity. 432
Furthermore, oil palm canopies are intersected by a dense network of roads used for 433
maintenance and harvest. Unsurprisingly, their near-infrared reflectance is high but very 434
homogeneous. Overall, spectral data density curves show considerable overlap across forest 435
disturbance levels, especially between forests that experienced moderate to high logging 436
pressures, and between heavily logged forests and mature oil palm plantations. 437
438
4.2 Modelling structure - sensor data relationships 439
Statistically fitting observed values of biophysical attributes, such as leaf area index, to 440
corresponding spectral vegetation indices is a common upscaling procedure. Spectral 441
vegetation indices exploit the differential between near infrared and red reflectance and the 442
ratio of these two spectral bands. This should render them more sensitive to the vegetation 443
component, while minimizing external effects caused by solar irradiance, atmospheric 444
conditions, and canopy background (Qi et al., 1995). Vegetation indices have consequently 445
been linked to productivity and canopy physical traits (Pfeifer et al., 2012; Sims et al., 2006; 446
Sjöström et al., 2011). Here, we show that vegetation greenness explains only 13 % of 447
biomass variability at plot level and heavily underestimates biomass in denser forests. Hence, 448
assuming vegetation greenness to be a direct surrogate for on-the-ground forest structure, 449
19
productivity or biodiversity (Pettorelli et al., 2005) is too simplistic, and is unlikely to 450
produce reliable outcomes for degradation landscapes in the humid tropics. 451
We demonstrate that spectral intensity data themselves can be used to estimate 452
biophysical structure, in particular for canopy attributes. By including texture information 453
into our mathematical models, we could significantly improve the models’ predictive 454
capacity, especially for canopy structure at stand level. Texture and spectral data combined 455
can be used to separate oil palm from moderately disturbed to undisturbed forests. The plots 456
measured at SAFE encompass the gradient of biomass found across the wider landscape, 457
setting this study apart from other tree plot studies (Marvin et al., 2014). Using tree census 458
data from these plots allowed us to map biomass, and thus carbon, across the landscape. 459
However, despite using data from more than 12 ha of surveyed area, we found that more 460
than 40% of the variation in plot attributes remain unexplained. These uncertainties in the 461
upscaling algorithm need to be accounted for when using the resulting maps for research and 462
conservation planning at landscape scales. Sources of observed uncertainties may include 463
inaccuracies in field measurements (Jonckheere, Muys, & Coppin, 2005; Phillips, Brown, 464
Schroeder, & Birdsey, 2000) or noise introduced through satellite image acquisition and 465
correction procedures (Dungan, 2003). Canopy reflectance at stand level is affected by many 466
factors operating at different spatial scales, i.e. transmittance of leaves, amount and 467
arrangement of leaves in the canopy, and characteristics of other canopy components. Large 468
time-spans between date of measurement in the field and date of image acquisition can 469
introduce errors, especially in seasonal environments. However, climate at SAFE landscape is 470
aseasonal and no dry months occurred between September 2011 and January 2013, the period 471
during which our data was collected. Canopy structure was measured in the field in the same 472
year (and for some plots within a month of image acquisition). Biomass was measured 473
approximately one year prior to image acquisition, which may contribute to errors in 474
20
upscaling algorithms as Borneo’s forests typically show high wood growth rates (9.73 Mg 475
dry mass per ha, Banin et al., 2014). However, these high growth rates are caused primarily 476
through tall trees of the Dipterocarpaceae family (Banin et al., 2014), which had been 477
targeted by selective logging rounds in disturbed forests. Low spectral resolution of the 478
images may limit our ability to distinguish between forest degradation stages at higher 479
disturbance levels, especially when analysing texture. 480
Spectral reflectance data are good indicators of rates (i.e. photosynthesis), but less 481
reliable indicators of vegetation state (e.g. leaf area index), due to non-linearity in 482
relationships (Sellers, 1987). In our study, we use empirical approaches to estimate 483
biophysical attributes from remote sensing data and to subsequently use the derived statistical 484
relationships to generate continuous surfaces of these attributes across the SAFE landscape. 485
A major drawback of this computationally efficient approach is that it is a site and sensor-486
specific one (Pfeifer et al., 2012). Physically based approaches using radiative transfer theory 487
model photon transport in vegetation canopies and can be adapted for various conditions 488
(Myneni et al., 2002). They are more powerful for understanding the mechanisms behind 489
observed variations in the spectral reflectance at the scale of leaves and individual trees, but 490
they are difficult to implement, scale- and vegetation type dependent and there is no universal 491
canopy radiative transfer model suitable for all canopies (Yin et al., 2015). 492
Complex patterns of topography and disturbance underlie spatial variation in above-493
ground carbon at landscape scale (Gale 2000, Pfeifer et al. 2015). Environmental factors 494
other than disturbance measured at SAFE probably affect above-ground live tree biomass, as 495
terrain steepness may affect both forest accessibility for logging but also forest growth. 496
Including micro-topographic drivers into models to map AGB variation across a forest 497
degradation landscape will be the subject of future studies. 498
499
Formatted: Font: 12 pt, Font color: Auto
21
4.3 Degradation impacts on forest structure at landscape scale 500
Up-scaled maps revealed a pronounced decline in aboveground live tree biomass with 501
increasing disturbance across the SAFE degradation landscape. These impacts are clearly 502
visible in both field data and maps, even a decade after logging. Harvesting rates computed 503
from these maps are higher than previously suggested in the literature (Struebig et al., 2013). 504
However, these were computed assuming that pre-logging biomass was similar to biomass in 505
nearby undisturbed forests, which may not necessarily be the case. In future analyses, 506
harvesting rates could be computed from time-series of biomass maps to avoid errors 507
introduced during such a space-for-time substitution approach. 508
Field data demonstrate a rapid recovery in forest canopy structure with the canopy 509
recovering to pre-disturbance levels a decade after logging. Yet, up-scaled maps, which cover 510
a larger survey area, suggest that both LAI and FCover are still reduced in logged compared 511
to primary forest stands (for details see Table 3 and Figure 5). Reduced LAI is linked to 512
higher daily maximum temperatures and lower daily humidity within tropical forests and at 513
this study site (Hardwick et al., 2015), and this in turn is likely to affect invertebrates and 514
invertebrate-mediated ecosystem functions (Ewers et al., 2015). We therefore suggest that 515
logging impacts on ecosystem processes and biodiversity are likely to be present more than a 516
decade after logging. Overall, our analyses suggest that both canopy structure and biomass 517
respond negatively to increased forest degradation. 518
Degradation impacts on forest structure appeared spatially heterogeneous, but not 519
necessarily in the way we would expect. For example, biomass was low in oil palms, and 520
quite consistently so. In contrast, AGB in primary forests varied strongly in space, despite 521
being higher on average when compared to disturbed forest stands. Canopy structure was 522
most variable in oil palms as we would expect given age differences of our oil palm stands, 523
but showed no trends with disturbance in forests, even though forest stands were increasingly 524
22
heterogeneous in their FCover response to disturbance. Natural tree falls causing localised 525
canopy damage may be one of a number of reasons explaining the heterogeneity observed in 526
the biophysical structure of primary forests. 527
528
4.4 Mapping High Carbon Stocks (HCS) in degraded landscapes 529
Current estimates suggest that at least one third of the world’s tropical humid forests are 530
affected by logging. Oil palm (Elais guineensis) is now grown across more than 13.5 million 531
ha of tropical humid lowland, often replacing high-biodiversity and high-biomass forest 532
(Fitzherbert et al., 2008). Malaysia and Indonesia, in particular, are major palm oil producers, 533
where demand for suitable agricultural land conflict with conservation initiatives (Koh & 534
Wilcove, 2007; Sodhi et al., 2004). In recent years, major conservation NGOs have been 535
working with corporate actors in palm oil and pulp & paper industries, developing a High 536
Carbon Stock approach to Forest Conservation Policies (Greenpeace International, 2013). 537
This approach aims to identify degraded lands suitable for development, cultivation and 538
expansion of oil palm plantations and to spatially delineate areas that should be conserved. 539
Our maps are sufficiently detailed to enable such decision-making and to identify (relatively) 540
high carbon stock areas within the wider study landscape. However, the practical value of 541
these maps will be dependent on the user’s ability to assess biomass variability in relative 542
rather than absolute terms. High biomass, little disturbed forests are clearly separated from 543
low biomass, heavily degraded forests and pockets of high-biomass forest are clearly visible 544
within the study landscape. Setting appropriate cut-off points for separating high from 545
moderate biomass forests is likely to be more challenging, given that the uncertainty in our 546
models and maps increases in high biomass zones. 547
548
5. Conclusions 549
23
Tropical forest landscapes are under increasing pressure worldwide as a result of 550
anthropogenic disturbance. Mapping forest degradation at near-real time and at the landscape 551
level can provide a valuable means of informing conservation actions or pre-emptive 552
management to limit losses. Our maps (Figure A3 in Appendix) depict the impacts of past 553
logging events on the canopy structure and biomass of tropical humid forests. These maps 554
can now be utilised to identify conservation win-wins (Strassburg et al., 2010), for example 555
by combining them with ongoing biodiversity surveys to delineate zones of high biodiversity 556
across forest types, and with measurements of carbon sequestration, hydrological cycles and 557
microclimate. These maps can also be manipulated to incorporate trade-offs between land use 558
demands and to take into account recent developments in fragmentation research. This 559
information ultimately feeds into High Carbon Stocks and High Conservation Value 560
approaches to forest and biodiversity conservation (Senior, Brown, Villalpando, & Hill, 561
2014). 562
563
Acknowledgement 564
We thank the Sime Darby Foundation, the Sabah Foundation, Benta Wawasan and the Sabah 565
Forestry Department for their generous support of the SAFE Project. We thank the Royal 566
Society’s South East Asia Rainforest Research Programme for logistical support in the field, 567
and Yayasan Sabah, Maliau Basin Management Committee, the State Secretary, Sabah Chief 568
Minister’s Departments, the Malaysian Economic Planning Unit and the Sabah Biodiversity 569
Council for permission to conduct research in Sabah. Data in the field were largely collected 570
with the help and support of the SAFE Project’s team of Research Assistants. MP and RME 571
were supported by European Research Council Project number 281986. We thank the 572
European Space Agency for providing RapidEyeTM under Category 1 - User Agreement to 573
MP (C1P 13735). 574
Deleted: 2575
24
576
References 577
Asner, G. P. (2009). Tropical forest carbon assessment: integrating satellite and airborne 578
mapping approaches. Environmental Research Letters, 4, doi:10.1088/1748-579
9326/4/3/034009. 580
Asner, G. P., Knapp, D. E., Broadbent, E. N., Oliveira, P. J. C., Keller, M., & Silva, J. N. 581
(2005). Selective Logging in the Brazilian Amazon. Science, 310, 480-482. 582
Asner, G. P., Powell, G. V. N., Mascaro, J., Knapp, D. E., Clark, J. K., et al. (2010). High-583
resolution forest carbon stocks and emissions in the Amazon. Proceedings of the 584
National Academy of Sciences of the United States of America, 107, 16738-16742. 585
Baccini, A., Goetz, S. J., Walker, W. S., Laporte, N. T., Sun, M., et al. (2012). Estimated 586
carbon dioxide emissions from tropical deforestation improved by carbon-density maps. 587
Nature Climate Change, 2, 182-185. 588
Banin, L., Lewis, S. L., Lopez-Gonzalez, G., et al. (2014). Tropical forest wood production: 589
A cross-continental comparison. Journal of Ecology, 102, 1025-1037. 590
Barton, K. (2014). MuMIn: Multi-model inference. R package version 1.10.0. From 591
http://cran.r-project.org/package=MuMIn 592
Berenguer, E., Ferreira, J., Gardner, T. A., Aragão, L. E. O. C., De Camargo, P. B., et al. 593
(2014). A large-scale field assessment of carbon stocks in human-modified tropical 594
forests. Global Change Biology, 20, 3713-3726. 595
Blaser, J., Sarre, A., Poore, D., & Johnson, S. (2011). Status of Tropical Forest Management 596
2011. ITTO Technical Series. doi:10.1017/S0032247400051135 597
Boyd, D. S. S., & Danson, F. M. M. (2005). Satellite remote sensing of forest resources: three 598
decades of research development. Progress in Physical Geography, 29, 1-26. 599
25
Burivalova, Z., Şekercioǧlu, Ç. H., & Koh, L. P. (2014). Thresholds of Logging Intensity to 600
Maintain Tropical Forest Biodiversity. Current Biology, 24, 1893-1898. 601
Castillo-Santiago, M. A., Ricker, M., & de Jong, B. H. J. (2010). Estimation of tropical forest 602
structure from SPOT-5 satellite images. International Journal of Remote Sensing, 31, 603
2767-2782. 604
Céline, E., Philippe, M., Astrid, V., Catherine, B., Musampa, C., & Pierre, D. (2013). 605
National forest cover change in Congo Basin: deforestation, reforestation, degradation 606
and regeneration for the years 1990, 2000 and 2005. Global Change Biology, 19, 1173-607
87. 608
Chander, G., Haque, M. O., Sampath, A., Brunn, A., Trosset, G., et al. (2013). Radiometric 609
and geometric assessment of data from the RapidEye constellation of satellites. 610
International Journal of Remote Sensing, 34, 5905-5925. 611
Chave, J., Réjou-Méchain, M., Búrquez, A., Chidumayo, E., Colgan, M. S., et al. (2014). 612
Improved allometric models to estimate the aboveground biomass of tropical trees. 613
Global Change Biology, 20, 3177-3190. 614
Didham, R. K., & Lawton, J. H. (1999). Edge Structure Determines the Magnitude of 615
Changes in Microclimate and Vegetation Structure in Tropical Forest Fragments1. 616
Biotropica, 31, 17-30. 617
Douglas, I. (1999). Hydrological investigations of forest disturbance and land cover impacts 618
in South-East Asia: a review. Philosophical Transactions of the Royal Society of London 619
B: Biological Sciences, 354, 1725-38. 620
Dungan, J. (2003). Spatial uncertainty in ecology: Implications for remote sensing and GIS 621
applications. International Journal of Geographical Information Science, 17, 93–94. 622
26
Edwards, D. P., Larsen, T. H., Docherty, T. D. S., Ansell, F. A., Hsu, Wet al. (2011). 623
Degraded lands worth protecting: the biological importance of Southeast Asia’s 624
repeatedly logged forests. Proceedings of the Royal Society B, 278, 82-90. 625
Ewers, R. M., Didham, R. K., Fahrig, L., Ferraz, G., Hector, A., et al. (2011). A large-scale 626
forest fragmentation experiment: the Stability of Altered Forest Ecosystems Project. 627
Philosophical Transactions of the Royal Society B: Biological Sciences, 366, 3292-628
3302. 629
Ewers, R. M., Boyle, M. J. W., Gleave, R., Plowman, N. S., Benedick, S., et al. (2015). 630
Logging cuts the functional dominance of social insects in tropical rainforest. Nature 631
Communications, 6, doi:10.1038/ncomms7836 632
Figueira, A. M. E. S., Miller, S. D., de Sousa, C. A. D., Menton, M. C., Maia, A. R., da 633
Rocha, H. R., & Goulden, M. L. (2008). Effects of selective logging on tropical forest 634
tree growth. Journal of Geophysical Research, 113, 1-11. 635
Fitzherbert, E. B., Struebig, M. J., Morel, A., Danielsen, F., Brühl, C. A., Donald, P. F., & 636
Phalan, B. (2008). How will oil palm expansion affect biodiversity? Trends in Ecology 637
& Evolution, 23, 538-545. 638
Frolking, S., Palace, M. W., Clark, D. B., Chambers, J. Q., Shugart, H. H., & Hurtt, G. C. 639
(2009). Forest disturbance and recovery: A general review in the context of spaceborne 640
remote sensing of impacts on aboveground biomass and canopy structure. Journal of 641
Geophysical Research: Biogeosciences, 114, doi:10.1029/2008JG000911 642
Gibson, L., Lee, T. M., Koh, L. P., Brook, B. W., Gardner, T. A., et al. (2011). Primary 643
forests are irreplaceable for sustaining tropical biodiversity. Nature, 478, 378-381. 644
Glenn, E. P., Huete, A. R., Nagler, P. L., & Nelson, S. G. (2008). Relationship Between 645
Remotely-sensed Vegetation Indices, Canopy Attributes and Plant Physiological 646
27
Processes: What Vegetation Indices Can and Cannot Tell Us About the Landscape. 647
Sensors, 8, 2136-2160. 648
Godard, E., Stoll, E., Anderson, C., Schulze, R., & D’Souza, B. (2012). Integrating Advanced 649
Calibration Techniques into Routine Spacecraft Operations. From 650
http://www.spaceops2012.org/proceedings/documents/id1275279-paper-001.pdf. The 651
12th International Conference on Space Operations. Last accessed 19/03/2015. 652
Greenpeace International. (2013). Identifying High Carbon Stock (HCS) Forest for 653
Protection. 654
http://www.greenpeace.org/international/Global/international/briefings/forests/2013/HC655
H-Briefing-2013.pdf. Last accessed 19/03/2015. 656
Grün, B., Kosmidis, I., & Zeileis, A. (2012). Extended Beta Regression in R: Shaken, Stirred, 657
Mixed, and Partitioned. Journal of Statistical Software, 48, 1-25. 658
Hansen, M. C., Potapov, P. V, Moore, R., Hancher, M., Turubanova, S., et al. (2013). High-659
resolution global maps of 21st-century forest cover change. Science, 342, 850-3. 660
Hardwick, S. R., Toumi, R., Pfeifer, M., Turner, E. C., Nilus, R., & Ewers, R. M. (2015). The 661
relationship between leaf area index and microclimate in tropical forest and oil palm 662
plantation: Forest disturbance drives changes in microclimate. Agricultural and Forest 663
Meteorology, 201, 187-195. 664
Harris, N. L., Brown, S., Hagen, S. C., Saatchi, S. S., Petrova, S., et al. (2012). Baseline Map 665
of Carbon Emissions from Deforestation in Tropical Regions. Science, 336, 1573-1576. 666
Hijmans, R. J., & van Etten, J. (2010). raster: Geographic analysis and modeling with raster 667
data. R Package Version. 668
Hugentobler, M. (2008). Quantum GIS. In Encyclopedia of GIS. 669
Jiang, Z., Huete, A. R., Didan, K., & Miura, T. (2008). Development of a two-band enhanced 670
vegetation index without a blue band. Remote Sensing of Environment, 112, 3833-3845. 671
28
Jonckheere, I. G., Muys, B., & Coppin, P. R. (2005). Derivative analysis for in situ high 672
dynamic range hemispherical photography and its application in forest stands. 673
Geoscience and Remote Sensing Letters, IEEE, 2, 296-300. 674
Joshi, N., Mitchard, E. T., Woo, N., Torres, J., Moll-Rocek, J., et al. (2015). Mapping 675
dynamics of deforestation and forest degradation in tropical forests using radar satellite 676
data. Environmental Research Letters, 10, doi:10.1088/1748-9326/10/3/034014 677
Koh, L. P., & Wilcove, D. S. (2007). Cashing in palm oil for conservation. Nature, 448, 993-678
994. 679
Koh, L. P., & Wilcove, D. S. (2008). Is oil palm agriculture really destroying tropical 680
biodiversity? Conservation Letters, 1, 60-64. 681
Lambin, E. F. (1999). Monitoring forest degradation in tropical regions by remote sensing: 682
some methodological issues. Global Ecology & Biogeography, 8, 191-198. 683
Laporte, N. T., Stabach, J. A., Grosch, R., Lin, T. S., & Goetz, S. J. (2007). Expansion of 684
industrial logging in Central Africa. Science, 316, 1451 685
Lucas, R. M., Clewley, D., Accad, A., Butler, D., Armston, J., et al. (2014). Mapping forest 686
growth and degradation stage in the Brigalow Belt Bioregion of Australia through 687
integration of ALOS PALSAR and Landsat-derived foliage projective cover data. 688
Remote Sensing of Environment, 155, 42-57 689
Marvin, D. C., Asner, G. P., Knapp, D. E., Anderson, C. B., Martin, R. E., et al. (2014). 690
Amazonian landscapes and the bias in field studies of forest structure and biomass. 691
Proceedings of the National Academy of Sciences of the United States of America, 111, 692
5224-5232. 693
Mermoz, S., Réjou-Méchain, M., Villard, L., Le Toan, T., Rossi, V., & Gourlet-Fleury, S. 694
(2015). Decrease of L-band SAR backscatter with biomass of dense forests. Remote 695
Sensing of Environment, 159, 307-317. 696
29
Myneni, R. B., Yang, W., Nemani, R. R., Huete, A. R., Dickinson, R. E., et al. (2007). Large 697
seasonal swings in leaf area of Amazon rainforests. Proceedings of the National 698
Academy of Sciences of the United States of America, 104, 4820-4823. 699
Naidoo, L., Mathieu, R., Main, R., Kleynhans, W., Wessels, K., Asner, G., & Leblon, B. 700
(2015). Savannah woody structure modelling and mapping using multi-frequency (X-, 701
C- and L-band) Synthetic Aperture Radar data. ISPRS Journal of Photogrammetry and 702
Remote Sensing, 105, 234-250. 703
Neteler, M., Bowman, M. H., Landa, M., & Metz, M. (2012). GRASS GIS: A multi-purpose 704
open source GIS. Environmental Modelling and Software, 31, 124-130. 705
Pettorelli, N., Vik, J. O., Mysterud, A., Gaillard, J.-M., Tucker, C. J., & Stenseth, N. C. 706
(2005). Using the satellite-derived NDVI to assess ecological responses to 707
environmental change. Trends in Ecology & Evolution, 20, 503-10. 708
Pfeifer, M., Gonsamo, A., Disney, M., Pellikka, P., & Marchant, R. (2012). Leaf area index 709
for biomes of the Eastern Arc Mountains: Landsat and SPOT observations along 710
precipitation and altitude gradients. Remote Sensing of Environment, 118, 103-115. 711
Pfeifer, M., Lefebvre, V., Gonsamo, A., Pellikka, P., Marchant, R., Denu, D., & Platts, P. 712
(2014). Validating and Linking the GIMMS Leaf Area Index (LAI3g) with 713
Environmental Controls in Tropical Africa. Remote Sensing, 6, 1973-1990. 714
Pfeifer, M., Lefebvre, V., Turner, E., Cusack, J., Khoo, M.S. (2015). Deadwood biomass: an 715
underestimated carbon stock in degraded tropical forests? Environmental Research 716
Letters, 10, 10.1088. 717
Phillips, D. L., Brown, S. L., Schroeder, P. E., & Birdsey, R. A. (2000). Toward error 718
analysis of large-scale forest carbon budgets. Global Ecology and Biogeography, 9, 305-719
313. 720
30
Pinard, M. A., & Putz, F. E. (1996). Retaining Forest Biomass by Reducing Logging 721
Damage. Biotropica, 28, 278-295. 722
QGIS Development Team. (2012). QGIS Geographic Information System. Open Source 723
Geospatial Foundation Project. http://qgis.osgeo.org. Qgisorg. Retrieved from 724
http://www.qgis.org/ 725
Qi, J., Chehbouni, A., Huete, A. R., Kerr, Y. H., & Sorooshian, S. (1994). A modified soil 726
adjusted vegetation index. Remote Sensing of Environment, 48, 119–126. 727
R Development Core Team. (2013). R Development Core Team. R: A Language and 728
Environment for Statistical Computing. 729
Reynolds, G., Payne, J., Sinun, W., Mosigil, G., & Walsh, R. P. D. (2011). Changes in forest 730
land use and management in Sabah, Malaysian Borneo, 1990-2010, with a focus on the 731
Danum Valley region. Philosophical Transactions of the Royal Society of London B: 732
Biological Sciences, 366, 3168-76. 733
Riou, R., & Seyler, F. (1997). Texture analysis of tropical rain forest infrared satellite images. 734
Photogrammetric Engeneering and Remote Sensing, 63, 515-521. 735
Robinson, C., Saatchi, S., Neumann, M., & Gillespie, T. (2013). Impacts of spatial variability 736
on aboveground biomass estimation from l-band radar in a temperate forest. Remote 737
Sensing, 5, 1001-1023 738
Rose, R. A., Byler, D., Eastman, J. R., Fleishman, E., Geller, G., et al. (2014). Ten Ways 739
Remote Sensing Can Contribute to Conservation. Conservation Biology, 740
doi:10.1111/cobi.12397 741
RSPO. (2013). Adoption of Principles and Criteria for the Production of Sustainable Palm 742
Oil (p. 70). From http://www.rspo.org/file/revisedPandC2013.pdf. Last accessed 743
19/03/2015. 744
31
DeFries, R., Asner G., Achard, F., Justtice, C., Laporte, N., et al. (2005). Monitoring 745
Tropical Deforestation for Emerging Carbon Markets. In: Tropical Deforestation and 746
Climate Change ( Moutinho, S. & Schwartzman S.). Amazon Institute for 747
Environmental Research. 748
http://publications.jrc.ec.europa.eu/repository/handle/JRC31742. Last accessed 749
19/03/2015. 750
Rutishauser, E., Noor’an, F., Laumonier, Y., Halperin, J., Hergoualc’h, K., & Verchot, L. 751
(2013). Generic allometric models including height best estimate forest biomass and 752
carbon stocks in Indonesia. Forest Ecology and Management, 307, 219-225. 753
Senior, M. J. M., Brown, E., Villalpando, P., & Hill, J. K. (2014). Increasing the Scientific 754
Evidence Base in the “High Conservation Value” (HCV) Approach for Biodiversity 755
Conservation in Managed Tropical Landscapes. Conservation Letters, Early View 756
online. doi:10.1111/conl.12148 757
Simas, A. B., Barreto-Souza, W., & Rocha, A. V. (2010). Improved estimators for a general 758
class of beta regression models. Computational Statistics and Data Analysis, 54, 348-759
366. doi:10.1016/j.csda.2009.08.017 760
Sims, D. A., Rahman, A. F., Cordova, V. D., El-Masri, B. Z., Baldocchi, D. D., et al. (2006). 761
On the use of MODIS EVI to assess gross primary productivity of North American 762
ecosystems. Journal of Geophysical Research: Biogeosciences, 111. 763
doi:10.1029/2006JG000162 764
Sjöström, M., Ardö, J., Arneth, A., Boulain, N., Cappelaere, B., et al. (2011). Exploring the 765
potential of MODIS EVI for modeling gross primary production across African 766
ecosystems. Remote Sensing of Environment, 115, 1081–1089. 767
32
Slik, J. W. F., Bernard, C. S., Breman, F. C., VAN Beek, M., Salim, A., & Sheil, D. (2008). 768
Wood density as a conservation tool: quantification of disturbance and identification of 769
conservation-priority areas in tropical forests. Conservation Biology, 22, 1299-308. 770
Slik, J. W. F., Paoli, G., Mcguire, K., Amaral, I., Barroso, J., et al. (2013). Large trees drive 771
forest aboveground biomass variation in moist lowland forests across the tropics. Global 772
Ecology and Biogeography, 22, 1261-1271. 773
Sodhi, N. S., Koh, L. P., Brook, B. W., & Ng, P. K. L. (2004). Southeast Asian biodiversity: 774
An impending disaster. Trends in Ecology and Evolution, 19, 654-660. 775
Strassburg, B. B., Kelly, A., Balmford, A., Davies, R. G., Gibbs, H. K., et al. (2010). Global 776
congruence of carbon storage and biodiversity in terrestrial ecosystems. Conservation 777
Letters, 3, 98-105. 778
Struebig, M. J., Turner, A., Giles, E., Lasmana, F., Tollington, S., et al. (2013). Quantifying 779
the biodiversity value of repeatedly logged rainforests: gradient and comparative 780
approaches from Borneo. Advances in Ecological Research, 48, 183-224. 781
Tropek, R., Sedláček, O., Beck, J., Keil, P., Musilová, Z., et al. (2014). Comment on “High-782
resolution global maps of 21st-century forest cover change”. Science, 344, 981. 783
Tucker, C. J. (1979). Red and photographic infrared linear combinations for monitoring 784
vegetation. Remote Sensing of Environment, 8, 127-150. 785
Vermote, E. F., Tanre, D., Deuze, J. L., Herman, M., & Morcette, J. J. (1997). Second 786
Simulation of the Satellite Signal in the Solar Spectrum, 6S: an overview. IEEE 787
Transactions on Geoscience and Remote Sensing, 35, 675-686. 788
Walsh, R. P., & Newbery, D. M. (1999). The ecoclimatology of Danum, Sabah, in the 789
context of the world’s rainforest regions, with particular reference to dry periods and 790
their impact. Philosophical Transactions of the Royal Society of London. Series B, 791
Biological Sciences, 354, 1869-83. 792
33
Wang, C., Qi, J., & Cochrane, M. (2005). Assessment of tropical forest degradation with 793
canopy fractional cover from Landsat ETM+ and IKONOS imagery. Earth Interactions, 794
9, 1-18. 795
Wearn, O. R., Reuman, D. C., & Ewers, R. M. (2012). Extinction Debt and Windows of 796
Conservation Opportunity in the Brazilian Amazon. Science, 337, 228-232. 797
Weiss, M., & Baret, F. (2010). CAN-EYE V6. 1 USER MANUAL. From 798
http://www6.paca.inra.fr/can-eye/Download. Last accessed 19/03/2015. 799
Wilcove, D. S., Giam, X., Edwards, D. P., Fisher, B., & Koh, L. P. (2013). Navjot’s 800
nightmare revisited: logging, agriculture, and biodiversity in Southeast Asia. Trends in 801
Ecology & Evolution, 28, 531-40. 802
Woodhouse, I. H., Mitchard, E. T. a, Brolly, M., Maniatis, D., & Ryan, C. M. (2012). Radar 803
backscatter is not a “direct measure” of forest biomass. Nature Climate Change, 2, 556-804
557. 805
Yin, G., Li, J., Liu, Q., Fan, W., Xu, B., Zeng, Y., & Zhao, J. (2015). Regional Leaf Area 806
Index Retrieval Based on Remote Sensing: The Role of Radiative Transfer Model 807
Selection. Remote Sensing, 7, 4604-4625 808
Zvoleff, A. (2014). Calculate textures from grey-level co-occurrence matrices (GLCMs) in R. 809
Retrieved from http://cran.r-project.org/web/packages/glcm/index.html 810
811
812 Formatted: Space After: 8 pt
34
Table 1 Global beta-logistic models relating forest structure attributes to RapidEyeTM 813
satellite sensor data at 135 plots. Models include intensity data for sensor bands 2 (GREEN), 814
3 (RED), 4 (RED_EDGE) and 5 (NEAR_IR), as well as their respective grey-level 815
dissimilarities (DissB2, DissB3, DissB4, DissB5). Structure attributes include (aboveground 816
biomass, AGB, leaf area index, LAI, and fractional vegetation cover, FCover) MSAVI2, 817
which is computed from red and near-infrared intensities, replaced both predictors in the 818
global model for AGB. The predicted estimates were scaled between 1e-6 and 65 (AGB), 1e-819
6 and 100 (FCover) and 1e-6 and 10 (LAI) to construct the models. See Table 2 for model-820
averaging results. DEV - Model fit computed as model-explained deviance. AIC - Akaike 821
Information Criterion. 822
Predicted Global Model DEV AIC
AGB
scaled
exp(MSAVI2) + GREEN + RED_EDGE + DissB2 + DissB3
+ DissB4 + DissB5
0.62 -457.9
LAI
scaled
DissB3 + DissB4 + DissB5 + RED + RED_EDGE +
NEAR_IR
0.52 -293.8
FCover
scaled
DissB3 + DissB4 + DissB5 + RED + RED_EDGE +
NEAR_IR
0.38 -143.9
823
Deleted: ¶824
Deleted: green825
Deleted: red826
Deleted: red-edge827
Deleted: near-infrared828
35
Table 2 Results of model-averaging carried out on three global beta-logistic regression 829
models relating forest structure attributes to satellite derived spectral and texture information 830
(see Table 1 for global models). The mathematical relationships between forest structure and 831
satellite sensor data were used to upscale forest structure across the SAFE landscape. Relative 832
variable importance (RVI) was computed during multi-model averaging. RVI = 1 if the 833
predictor was present in all models. Coefficient estimates in bold: RVI = 1. DEV - Model fit 834
computed as model-explained deviance. N - Number of models computed during model-835
averaging. For further explanation of variables used see Table 1. 836
Predicted Model coefficients (with shrinkage) Phi DEV N
AGB
scaled
19.45a - 5.01*exp(MSAVI2a) - 2.39*GREENa +
1.08*RED_EDGEa + 2.65*DissB2a - 0.28*DissB3 -
0.13*DissB5 + 0.09*DissB4
12.15a 0.61 8
LAI
scaled
0.90c - 0.59*REDa + 0.41*RED_EDGEa -
0.11*NEAR_IRa - 0.53*DissB3d+ 1.08*DissB4a -
0.36*DissB5a
37.43a 0.52 2
FCover
scaled
2.66b - 0.66*REDb + 0.30*RED_EDGEd -
0.08*NEAR_IRc + 1.48*DissB4a - 0.42*DissB5c -
0.17*DissB3
8.93a 0.38 7
a P < 0.001, b P < 0.01, c P < 0.05, d P < 0.01 837
838
839
840
841
Deleted: ¶842
Deleted: green843
Deleted: red-edge844
Deleted: red845
Deleted: red-edge846
Deleted: near-infrared847
Deleted: red848
Deleted: red-edge849
Deleted: near-infrared850
Formatted: Space After: 8 pt
36
Table 3 Summary statistics for forest structure variation across disturbance levels using maps 851
derived from upscaling forest structure. Aboveground biomass (AGB): the distribution of 852
estimates was least variable in oil palms (F test, P < 0.001); AGB was more variable in 853
primary and lightly logged forest compared to twice-logged and salvage logged forest, but 854
less variable in twice compared to salvage logged forest (all P < 0.001). Leaf area index 855
(LAI): the distribution of estimates was most variable in oil palm; LAI was more variable in 856
primary forest compared to other forests, but less variable in lightly versus twice-logged 857
forest, and less variable in twice versus salvage logged forest (all P < 0.001). Fractional 858
vegetation cover (FCover): the distribution of estimates was most variable in oil palm; 859
FCover was less variable in lightly logged versus other forest types, less variable in primary 860
compared to twice-logged forest and less variable in twice-logged versus salvage logged 861
forest (all P < 0.001). 862
863
AGB (t/ha) LAI (m2/m2) FCover (%)
Mean Median SD Mean Median SD Mean Median SD
Oil Palm 38.1 33.2 25.2 2.47 2.48 0.7 56.23 57.85 10.5
Salvage L 95.4 75.4 72.8 3.78 3.84 0.5 70.96 71.13 6.9
Twice L 122.4 112.6 52.6 3.82 3.81 0.4 74.32 75.04 4.8
Lightly L 163.2 123.2 102.7 4.12 4.10 0.3 74.71 74.10 3.5
Primary 350.4 337.5 111.9 4.43 4.46 0.5 81.03 81.43 4.1
864
865
866
867
Deleted: ¶868
Formatted: Superscript
Formatted: Superscript
37
Fig. 1 Location of vegetation plots within the SAFE landscape. The average altitude of plots 869
is 439 m above sea level [range: 238 - 618 m] based on ASTER global digital elevation 870
model 2 (ASTER GDEM is a product of METI and NASA). Selected close-ups of the 871
sampling design in forest stands and oil palm plantations are overlaid on False Colour 872
Composites of two RapidEyeTM images. 873
874
Fig. 2 Forest attributes aggregated at the level of forest types. (A) Above-ground biomass in 875
live tree (AGB) decreased significantly from unlogged, protected forest stands (OG) to 876
severely degraded stands (Salvage logged) and oil palm plantations (OP). (B) and (C) 877
Canopy structure (Leaf Area Index, LAI; Fraction of vegetation cover, FCover) does not 878
differ significantly between forest types, although it does differ markedly between forest 879
stands and OP. For details on pairwise significant differences see text. 880
881
Fig. 3 Spectral and textural image data derived from RapidEyeTM imagery for plots and 882
aggregated to the level of forest types. Oil palm blocks differed significantly from forests in 883
all satellite derived information. (A) Red intensity was lower in primary and lightly logged 884
forest compared to salvage logged and twice-logged forest (P < 0.05). (B) Vegetation 885
greenness expressed as MSAVI2 was higher in primary compared to twice-logged and 886
salvage logged forest (P < 0.001) and lower in twice-logged versus salvage logged forest (P 887
< 0.005). (C) Red dissimilarity was higher in primary and twice-logged forest compared to 888
salvage logged forest (P < 0.001). Primary forest had significantly lower near-infrared band 889
dissimilarity compared to twice-logged forest (P < 0.05) but significantly higher near-infrared 890
band dissimilarity compared to salvage logged forest (P < 0.001). (D) Near-infrared 891
dissimilarity was lower in salvage logged forest compared to twice-logged (P < 0.0001) and 892
lightly logged forest (P < 0.001). 893
Deleted: ¶894
38
895
Fig. 4 Single-predictor relationships between forest attributes and spectral data derived from 896
RapidEyeTM imagery at plot level and aggregated to stand level. At stand level, aboveground 897
biomass increased exponentially with increasing vegetation greenness. Leaf area index and 898
fractional vegetation cover decreased linearly with increasing red intensity. At plot level, 899
relationships are weaker but still significant. For details on models and models fits see text. 900
901
Fig. 5 Density curves computed from histograms of structure attributes and extracted for each 902
disturbance level from the newly generated maps. For details on the extraction see text. All 903
density curves differed significantly from each other (two-sample Kolmogorov-Smirnov test; 904
P < 0.001). All forest types differ significantly from each other in their structure. On average, 905
AGB (t / ha), LAI and FCover decreased sequentially from primary forest to lightly logged, 906
twice-logged and salvage logged forests (see also Table 3). 907
908
Fig. 6 Derived maps of forest structure attributes (above-ground biomass in live tree, AGB, 909
leaf area index, LAI, and fractional vegetation cover, FCover) at the SAFE degradation 910
landscape. Details on models used to generate these maps are provided in Table 2. 911
39
912
913
Figure 1 Location of vegetation plots within the SAFE landscape. 914
40
915
916
Figure 2 Forest attributes aggregated at the level of forest types.917
41
918
919
Figure 3 Spectral and textural image data derived from RapidEyeTM imagery for plots and 920
aggregated to the level of forest types. Red intensity is given in µW/(cm2 * sr * nm). 921
922
923
Formatted: Superscript
42
924
925
Figure 4 Single-predictor relationships between forest attributes and spectral data derived 926
from RapidEyeTM imagery at plot level and aggregated to stand level.927
43
928
929
930
Figure 5 Density curves computed from histograms of structure attributes and extracted for 931
each disturbance level from the newly generated maps.932
44
933
934
Figure 6 Derived maps of forest structure attributes (above-ground biomass in live tree, 935
AGB, leaf area index, LAI, and fractional vegetation cover, FCover) at the SAFE degradation 936
landscape. 937
938 Formatted: Left, Space After: 8 pt, Line spacing: Multiple 1.08 li
45
Appendix 939
940
Table A1 Details on RapidEyeTM imagery used in this study. There were 193 vegetation plots 941
(25 m x 25 m) established at SAFE in 2010 and distributed among 17 sampling blocks (table 942
1): three oil palm plantation blocks (OP1 and OP2 planted in 2006 and OP3 planted in 2000), 943
two primary forest blocks (OG1 and OG2), two lightly or illegally logged forest blocks (OG3 944
and VJR), four twice-logged forest blocks (LFE, LF1 - LF3) and six salvage-logged forest 945
blocks that are currently being converted into a fragmented agricultural landscape (A - F) 946
(Ewers et al 2011). We thank the European Space Agency for providing RapidEyeTM under 947
Category 1 - User Agreement to Marion Pfeifer (C1P 13735). Res - Resolution, IA - 948
Incidence angle, SE - Illumination elevation angle, SVA - Spacecraft viewing angle. 949
950
Tile Date Blocks (forests and
oil palms) covered
Res IA SE SVA
5041214 21/09/2012 OG1 to OG 3 5 m 18.56 80.84 16.79
5041215 21/09/2012 - 5 m 16.37 81.04 16.79
5041216 13/05/2012 - 5 m 16.78 75.08 -13.16
5041217 18/08/2012 F, VJR, LFE, B, C,
D, E, OP1, OP2, OP3
5 m 7.20 78.77 -6.29
5041218 24/01/2013 A, LF1 to LF3 5 m 6.53 63.36 6.78
951
952
Deleted: ¶953
46
Table A2 Pearson’s product moment correlation coefficients (Spearman’s rho) between 954
predictor variables (texture indices: mean, variance contrast, dissimilarity; spectral data) 955
computed for bands 2 to 5 (B2 to B5) and extracted for the field derived point data. ns - 956
correlation not significant at P = 0.05. Showing results for window size 9 pixels x 9 pixels 957
only. B2 - GREEN, B3 - RED, B4 - RED_EDGE, B5 - NEAR_IR. 958
959
Mean Variance Contrast Dissimilarity
B2 B3 B4 B5 B2 B3 B4 B5 B2 B3 B4 B5 B2 B3 B4 B5
Mean
B2 - .88 .89 .75 .34 .26 .39 .41 .39 .25 .23 .24 .46 .28 .25 .28
B3 - - .54 .31 - .56 .44 .20 - .53 .24 .10 - .57 .22 .13
B4 - - - .92 - - .36 .56 - - .29 .42 - - .33 .45
B5 - - - - - - - .55 - - - .38 - - - .40
Vari
an
ce
B2 - - - - - .96 .79 .16 .62 .57 .34 .09 .56 .55 .26 .08
B3 - - - - - - .68 .02 - .64 .25 -.02 - .64 .16 -
.03
B4 - - - - - - - .62 - - .65 .51 - - .62 .49
B5 - - - - - - - - - - - .86 - - - .85
Con
trast
B2 - - - - - - - - - .85 .58 .14 .94 .71 .48 .14
B3 - - - - - - - - - - .41 -.1 - .91 .30 -
.06
B4 - - - - - - - - - - - .74 - - .96 .69
B5 - - - - - - - - - - - - - - - .98
Dis
sim
ilari
ty
B2 - - - - - - - - - - - - - .61 .48 .20
B3 - - - - - - - - - - - - - - .22 -
.09
B4 - - - - - - - - - - - - - - - .74
B5 - - - - - - - - - - - - - - - -
Sp
ectr
al
B2 .45 - - - .37 - - - .50 - - - .41 - - -
B3 .32 .70 - - .45 .56 - - .56 .73 - - .44 .74 - -
B4 .47 .49 .31 - .19 .23 -.03 - .30 .37 -.11 - .25 .41 -.16 -
B5 .25 .04 .28 .41 -.00 .00 -.14 -.19 -.01 .04 -.29 -.27 -.00 .10 -.32 -
.27
960
961
Deleted: 1962
Deleted: 1963
Formatted: Left, Space After: 8 pt
47
Table A3 Models fits (R2adj ) of linear single - predictor models relating texture indices to 964
biophysical attributes at plot level. ns - correlation not significant at P = 0.05. AGB - Live 965
tree aboveground biomass, Fcov - Fractional vegetation cover. LAI - Leaf area index. B2 - 966
GREEN, B3 - RED, B4 - RED_EDGE, B5 - NEAR_IR. 967
968
Window 3 x 3 Window 5 x 5 Window 7 x 7 Window 9 x 9
AGB LAI FCov AGB LAI FCov AGB LAI FCov AGB LAI FCov
Mean
B2 .16 ns ns .17 ns ns .18 ns ns .19 ns ns
B3 ns .15 .10 .02 .16 .11 .02 .17 .12 .03 .17 .13
B4 .32 ns ns .33 ns ns .34 ns ns .34 ns ns
B5 .42 .04 .03 .43 .04 .03 .44 .04 .03 .43 .03 .03
Vari
an
ce
B2 ns .06 .-3 ns .09 .05 ns .10 .06 ns .11 .07
B3 ns .17 .07 ns .19 .09 ns .19 .11 ns .19 .11
B4 .02 ns ns .04 ns ns .04 ns ns .02 ns ns
B5 .16 .07 .04 .21 .06 .03 .22 .05 ns .19 .03 ns
Con
trast
B2 ns .02 ns ns .03 ns ns .05 ns ns .07 ns
B3 ns .08 ns ns .10 .02 ns .13 .04 .14 .18 .06
B4 ns ns ns ns ns ns ns ns ns ns ns ns
B5 .14 ns ns .15 ns ns .14 ns ns .11 ns ns
Dis
sim
ilari
ty
B2 ns ns ns ns ns ns ns .03 ns ns .04 ns
B3 .03 .19 .09 .04 .21 .10 .05 .25 .13 .06 .28 .15
B4 .04 ns .02 .05 ns .03 .05 ns .02 .04 ns ns
B5 .16 ns ns .17 ns ns .17 ns ns .14 ns ns
969
970
971
972
Deleted: ¶973
48
974
975 Figure A1 Spectral vegetation indices derived from RapidEyeTM imagery for plots and 976
aggregated to the level of forest types. We calculated the Normalized Difference Vegetation 977
Index as NDVI = (Near_IR - RED)/(Near_IR + RED), the Enhanced Vegetation Index as 978
EVI = 2.5 * (NEAR_IR - RED)/( NEAR_IR + 2.4 * RED + 1), the Modified Soil-Adjusted 979
Vegetation Index as MSAVI2 = ((2 * NEAR_IR + 1) - (( 2 * NEAR_IR + 1 )^{2} - 8 * 980
(NEAR_IR - RED))^{1/2}) / 2, and the Red Edge Index as REI = ((NEAR_IR - RED)/( 981
NEAR_IR + RED)) +1 Spectral vegetation indices were highly correlated with each other 982
across plots as indicated by Pearson’s product-moment correlation, r (NDVI ~ MSAVI2: r = 983
0.996, NDVI ~ EVI2: r = 0.994, EVI2 ~ MSAVI2: r = 0.985, MSAVI2 ~ REI: r = 0.882, 984
NDVI ~ REI: r = 0.849, EVI2 ~ REI: r = 0.872). 985
986
987
Deleted: ¶988
Deleted: ¶989
Deleted: were 990
Deleted: ¶991 ¶992
Formatted: Indent: Left: 0 cm, First line: 0 cm
49
993
Figure A2 Comparison of predicted forest attributes (using our final models) against forest 994
attributes observed in the field at the level of plots (grey dots) and at the level of forest types 995
(mean ±standard deviation). 996
997
Formatted: Space After: 0 pt, Line spacing: Double
Formatted: Font: Not Bold
50
998
Top Related