Mapping dynamics of deforestation and forest degradation in tropical forests using radar satellite data

Surface moisture can vary backscatter and decrease its contrast over forests and bare-ground. To minimize the problem, two dry-season scenes were acquired from the phased array L-band Synthetic Aperture Radar sensor aboard the Advanced Land Observing Satellite (ALOS PALSAR) in July–August for each year from 2007 to 2010 (scene IDs provided in table S1). Since terrain can significantly impact image projection and backscatter, scenes were terrain-corrected and radiometrically calibrated using the 90 m resolution Shuttle Radar Topography Mission dataset (Jarvis et al 2008). They were then converted to backscatter (co-polarized $\sigma _{{\rm HH}}^{0}$ and cross-polarized $\sigma _{{\rm HV}}^{0}$) using coefficients of Shimada et al (2009) and mapped at 15 m resolution in UTM projection (schematic processing chain is shown in figure 2(a)).

Zoom In Zoom Out Reset image size

Speckle, which is inherent in radar images (Woodhouse 2006a), may be problematic for accurate disturbance detection. Our images were moderately despeckled using the Enhanced Lee Filter (Lopes et al 1990) with a 3 × 3 pixel window to retain textural information. Image pixels were then averaged to 30 m resolution as a compromise to reduce speckle, but allow the detection of small-area disturbances and comparability to Landsat-based deforestation datasets (e.g. Asner et al 2010, Hansen et al 2013).

To reduce any remaining variability in backscatter unrelated to anthropogenic disturbances, images were calibrated to the base year (2007) by adjusting pixel values by the difference in average backscatter over a 510 × 510 m window (figures 2(b)–(c)). This procedure was chosen since it showed reasonable adjustment over known forest/non-forest areas (collected field data described in section 3.2.2), and corrected for variations over different land cover types locally. The chosen window size was much larger than most known disturbances in the region, and they were hence unlikely to be lost by the calibration. To ensure that large-area disturbances were not lost, backscatter was not adjusted if averages differed by more than a cut-off threshold; here, half the standard deviation of 2007 backscatter, chosen empirically based on field knowledge of large-area disturbances. The procedure showed no significant impact on the area-dependent analysis presented in section 4.4.

3.2.1. Detection algorithm

A preliminary investigation of our study region revealed frequently disturbed areas with a continuum of surrounding disturbed land. To maximize the detection of these areas, we exploited the advantage of multi-temporal and spatially continuous radar images. A time-series allows the same pixel to be observed multiple times, and hence allows more confidence in determining its status (disturbed/undisturbed), as compared to only two observations. To minimize detecting backscatter variations unrelated to anthropogenic disturbances, we used a local moving-window filtering procedure that assumes that disturbances are more likely to be real if they neighbour other disturbances. The designed algorithm hence (i) uses multi-temporal data as intermediate information to verify disturbance locations, but does not temporally categorize change pixels initially, and (ii) identifies areas with slow-recovering and fast-recovering backscatter change (i.e. areas where backscatter remained consistently changed compared to year 2007 for a period of two years or more, and for a period of one year, respectively) in close proximity, but does not spatially categorize these into deforested/degraded.

Areas known to have <30% forest cover in year 2000 (Hansen et al 2013), $\sigma _{{\rm HV}}^{0}\;\lt \;$−14.0 dB or Radar Forest Degradation Index >0.5 (figure 2(a)) (Mitchard et al 2012) in 2007 were discarded, removing bare-ground, infrastructure, water and some inundated swamps (4.7% of study area) from the analysis. Sucessional forest dynamics are not studied for these areas. Backscatter, particularly $\sigma _{{\rm HV}}^{0}$, was found to be related to biomass for the remaining areas (figure S3) and observed to decrease upon disturbance. Since forests tend to depolarize the radar signal (Woodhouse 2006a), giving high $\sigma _{{\rm HV}}^{0}$ values that reduce upon forest cover loss, our methodology uses the annual loss in $\sigma _{{\rm HV}}^{0}$ for disturbance detection.

To identify a disturbed pixel, the algorithm uses a set of thresholds (T1 to T5, table 1), explained in figure 3 and exemplified in figure 4. Five backscatter change maps ($\Delta \sigma _{{\rm HV}}^{0}\ 2010$–$2007$, $\Delta \sigma _{{\rm HV}}^{0}\ 2009$–$2007$, $\Delta \sigma _{{\rm HV}}^{0}\ 2008$–$2007$, $\Delta \sigma _{{\rm HV}}^{0}\ 2009$–$2008$ and $\Delta \sigma _{{\rm HV}}^{0}\ 2010$–$2009$ ) were produced and disturbance captured in two stages. First, slow-recovering changes are selected by eliminating pixels with change values that do not meet threshold T1 and are not seen consistently over two years or more (box A, output 1), and are spatially isolated from pixels detected in all years' slow-recovering change maps (output 2). Second, fast-recovering changes are selected by eliminating pixels with change values that do not meet threshold T1 (box B, output 3), and eliminating pixels that are spatially isolated from all years' fast-recovering or slow-recovering change maps (output 4). Locations of disturbances are then assigned to time periods to obtain annual disturbance maps (box C, output 5).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Table 1.  Thresholds selected for forest disturbance detection in the study area for the algorithm presented in figure 3. Thresholds were chosen empirically through a manual process, aimed at maximizing detecting known disturbances and minimizing false detections based on prior knowledge of the study area.

Threshold	Chosen value
T1	−1.5 dB
T2	56%
T3	2 pixels
T4	67.5%
T5	3 pixels

3.2.2. Training and validating detections

3.2.3. Estimating false positives due to speckle

Significant differences in the distribution of backscatter for disturbed and undisturbed, and slow- and fast-recovering, areas were recorded in the study region (figures 5(a)–(b)). Over the three year period, 2.3% (17 722 ha) of the study area was disturbed (rate of 0.78% yr⁻¹), with 1.3% under slow-recovery and 1.0% under fast-recovery. Of the disturbed land, 2,555 ha was found to be repeatedly disturbed during the three years, and there was decline in total disturbed area with each year (table 2). Disturbances were observed mostly along the IOH and secondary roads, although fast-recovering disturbances were more widespread (figure 5(c)).

Zoom In Zoom Out Reset image size

The proportion of change type (fast- and slow-recovering) differed over areas with different authorized land uses (figure 5(d)). Protected areas, conservation and Brazil nut concessions had the lowest percent of forest disturbance, and a higher fraction of fast-recovering changes, in comparison to mining concessions and areas under no allocated land use (e.g. lands in long-term use for mainly agricultural and pastoral farming).

Validation of the detection algorithm is reported using the sources introduced in section 3.2.2:

• (1)  
  Disturbances were detected in 1721 of 2740 pixels on the delineated agricultural and pastoral farms, a hit rate of 63% and miss rate of 37%, over the three years.
• (2)  
  Over the remote forest area, the difference in backscatter was 0.015 ± 1.477 dB (mean ±standard deviation) between 2007 and 2010 (figure S5). A false positive rate of 0.3% of area was detected (consistently 0.1% yr⁻¹). Of these, slow-recovery false positives were 0.04% of area, suggesting that there is more uncertainty associated with fast-recovering disturbances. Sensitivity of detections to simulated speckle (section 3.2.3) showed that only a small proportion of pixels, 0.04% yr⁻¹, are detected by chance. This suggests that at least some false positives in the remote area are genuine disturbances.
• (3)  
  Of the 250 Brazil nut concessions that reported logging, disturbances were detected in 212 concessions. For these, regression analysis between extracted volumes normalized by concessions' area and percentage of detected disturbed area showed a very weak but significant relation, particularly for the last year of data (r² = 0.15, p < 0.001) (figure S6). Disturbances are not expected to result from logging activities alone and log volumes do not necessarily have a high correlation with area logged. However, the result suggests that the area of detections picked up by radar are realistic to a significant extent.
• (4)  
  Of the 24 GPS-marked areas where logging permission was requested, disturbances were observed within 12, and in 15 when allowing for a 90 m buffer (threshold T5). The probability of these 15 being observed by chance is 0.04 or 4% (i.e. in 40 of 1000 iterations).
• (5)  
  Of the 34 felling gaps however, only 6 were detected within 60–90 m buffer with a probability of chance observation of 0.01 or 1% (i.e in 10 of 1000 iterations). The low detection rate is expected, since logging is sparse in the region and felling-gaps are smaller than the radar image resolution (Moll-Rocek et al 2014).

The sensitivity analysis to thresholds T1–T5 showed that the algorithm was most sensitive to T2 and T4, i.e. the fraction of pixels that must be disturbed within the neighbourhood of a pixel for it to be selected as disturbed, followed by T3 and T5, i.e. the neighbourhood size. A 0% false positive rate was achievable, although at the expense of reducing the hit rate to between 50 and 52% (figure S7).

To analyse the spatial patterns of detections, disturbances were visualized as forest disturbance 'events' (FDE) by grouping and converting neighbouring disturbed pixels to polygons. These polygons indicated that 85% of slow-recovery events were detected with fast-recovery events along edges, visibly related to the same land use. Fast-recovery events were more frequent (2.6 times) than and more isolated from slow-recovery events (70% neighboured slow-recovery events). Both types of disturbances were also common at the frontiers of already-cleared land (figures 4 and S8). It was noted that spatially delineating disturbed areas with abrupt boundaries risks ignoring surrounding lands in transitional disturbance. Hence, we fused neighbouring fast- and slow-recovery events as belonging to the same FDE, merging neighbouring deforestation and degradation and giving a continuous spatial gradient of disturbance.

To understand successional forest dynamics, we analysed fused FDEs occurring in the 2007–2008 period that fall into different land cover/use categories, delineated using an independent dataset in 2012 (figure 1(a)) (IIAP 2013, BAM 2013). Analysis of variance (ANOVA) tests revealed that events resulting in pastures, typically clearing full forest cover, were characterized with significantly lower backscatter prior to disturbance as compared to (unclassified) events in intact forests (difference of 1.1 dB, p < 0.001) and gold-mining (difference of 0.8 dB, p < 0.001) (figure 6). Such a result is expected, as pastures are concentrated along areas already under human-use (figure 1(a)) and gold-mining involves clearing relatively intact forests. Backscatter remains low in successive years following FDE for gold-mining as compared to increases observed for events in intact and secondary forests (difference of 1.5 and 1.2 dB respectively, p < 0.001), suggesting that the recovery of forest cover or biomass in gold-mining areas is significantly slower (test statistics described in tables S2–S4).

Zoom In Zoom Out Reset image size

Initial and recovered backscatter values were also related to the size of land cleared for each land use (examples in figure 7). Regression analysis revealed that (i) large-sized events for pastures and gold-mining begin with low backscatter in the year before clearance (slope = −0.51 and −0.47 respectively, p < 0.001), suggesting that the clearing of large areas is more common in lands with low biomass. Although a similar trend is visible in intact forests, the overall trend across events of different sizes is not significant (p > 0.5); (ii) large-sized events also have a large drop in backscatter during a FDE, particularly for gold mining (slope = −0.96, p < 0.001). The magnitude of change (ranging ∼3–5 dB) suggests that these large-areas are more likely to be fully cleared or with very low biomass immediately after disturbance. In contrast, small-sized events are more likely to be diffuse disturbances not involving full clearance; and (iii) large-sized events for pastures, and marginally for gold-mining (slope = 0.37, p = 0.02), recover backscatter faster in the years after FDE. Personal field observations (Woo 2012) recorded that parts of large pastures were commonly abandoned due to maintenance challenges. Small changes in biomass may cause large changes in backscatter when initial biomass values are low, as seen theoretically (Woodhouse 2006b, Brolly and Woodhouse 2012) and empirically (non-linear biomass-backscatter curve, figure S3). The result suggests that rapid cover and biomass accumulation post-disturbance are crucial successional dynamics in these lands.

Zoom In Zoom Out Reset image size

Our study presents a new algorithm to detect forest disturbances using radar, relying on a temporal trajectory of backscatter change and identifying areas with both slow-recovering (backscatter drop sustained for two or more years) and fast-recovering (backscatter drop sustained for one year) disturbances. In the rainforests of Madre de Dios, we estimated that 2.3% of the study area was disturbed between July 2007–2010 with a false-positive rate of 0.3% of area, and with disturbances most widespread during July 2007–August 2009. This peak is similar to that noted in previous studies (Asner et al 2010, Hansen et al 2013), and has been linked to the boom of international gold prices in 2008–2009 and increasing land use by migrants (Fraser 2009, Swenson et al 2011, Vuohelainen et al 2012, Asner et al 2013). Similar to previous studies, protected, conservation or sustainably managed areas were least disturbed, either due to the inhibitory effects of land concessions on deforestation or their remote location (Nunes et al 2012, Vuohelainen et al 2012).

Comparisons to optical-based deforestation datasets of Asner et al (2010) and Hansen et al (2013) however indicated that the total area of disturbance detected by radar, summing fast- and slow-recovering areas, is over twice their estimates (after removing pixels overlapping detections occurring before 2007 in both previous studies) (figure 8(a)). This is not necessarily a like-for-like comparison, as the studies demarcate annual forest loss and may exclude the observed extended gradient of disturbances surrounding affected areas (examples in figure S8) (although Asner et al (2010) includes 'disturbances', there are diffuse thinning from other land uses). We allowed for ±1 year buffer in the temporal classification of forest disturbance events and assessed the number of events that overlapped between studies by at least 25% of area. Most differences were for small-sized events, i.e. <30% of previous studies' events of 0.3 ha and 70% at 1.7 ha were detected in our map (figure 8 (b)). At ∼4 ha the Hansen et al (2013) study captured 50% and the Asner et al (2010) study captured 30% of our events (figures 8 (c)–(d)).

Zoom In Zoom Out Reset image size

The comparability of the performance of the datasets is limited, since we only assess end-products of optical-derived regional to global deforestation and our study targets mapping accurate provincial-level disturbances backed by detailed field-surveys instead. However, long-wavelength radar is expected to detect larger areas of disturbance, including degradation, mainly due to its sensitivity to changes within and below the forest canopy. Relations between disturbed area and timber-extract volumes and locations suggest that degradation is picked up to a significant extent in Madre de Dios. Some studies have relied on a combination of data, such as optical and airborne/satellite lidar (e.g. Asner et al 2010, Zhuravleva et al 2013), to specifically capture degradation. We hence urge the development of methods that integrate radar in such studies to aid frequent disturbance monitoring. However, constraints still remain on the availability, extent and costs of long-wavelength radar imagery, which may be overcome with products from ALOS-2 (JAXA 2014) and the future SAOCOM (CONAE 2014) and BIOMASS (LeToan et al 2011) missions.

Various scattering processes associated with land disturbances have shown to cause both decreases and increases in backscatter (Whittle et al 2012). In our study area, an examination of known disturbances showed that a loss in the cross-polarized signal, $\sigma _{{\rm HV}}^{0}$, dominated over disturbed lands. This is expected, since processes that may increase backscatter and cause large changes in the co-polarized signal (e.g. creating and maintaining hard forest edges for ranches) were not common practices. Further, backscatter may have low sensitivity to low-magnitude biomass losses in areas with high biomass due to apparent signal saturation (typically 100 Mg ha⁻¹ for L-band radar). However, significant detections including selective logging were picked up in such areas. This may be because large disturbances in vegetation structure that were relevant at the 30 m resolution (e.g. removal of a large tree), and to which backscatter is fundamentally related (Woodhouse et al 2012), accompanied these activities. Broadly, further research on backscatter–vegetation interactions and potential errors from saturation would greatly benefit the use of radar for disturbance detection over other study areas.

Limitations also lie in needing to set cut-off thresholds T1–T5 (table 1) in the algorithm. The algorithm performance was most sensitive to the spatial extent of disturbances. The thresholds inherently coarsen the output map 'resolution', losing very small-sized or slow biomass losses, and resulting in a miss rate of known disturbances of 37% and low detection rate of gaps from sparse selective logging. However, the algorithm is designed using a sequence of observations to ensure confidence in results. Inserting 'time-slices' of more frequent image acquisitions, such as those from the 14-day repeat cycle of ALOS-2, can allow for more flexibility in the use of thresholds. Nevertheless, environmental conditions such as soil moisture, changes in understory vegetation, sensor calibration drift and speckle are expected to contribute to noise and false positives, requiring local-calibration and ground-data to ensure accuracy.

Since deforestation and forest degradation may not constitute simple land cover conversions between two time periods, the radar-based change detection algorithm improves the characterization of these processes by allowing for fluctuations and reversibility in conversion, and the detection of repeated disturbances. A significant proportion of fast-recovering disturbance events (70%) neighboured slow-recovering events. This result highlights both the continuous gradient of land cover change (Guariguata et al 2009) and the possibility of degradation preceding and accompanying deforestation. While a recent synthesis of national REDD+ readiness revealed that most countries' direct intervention plans focus on reducing forest degradation (Salvini et al 2014), methods to quantify degradation using common remote sensing time-series are not well established (De Sy et al 2012). The technique described in this study may hence be of benefit, particularly in areas prone to subtle changes in forest cover, and where compliance with interventions requires frequent monitoring. Quantifying the type of disturbance (fast- and slow-recovery) can also aid characterizing the dynamics of land cover changes associated with legally granted land uses.

The importance of studying non-forest lands to assess the effectiveness of policies (Salvini et al 2014) and monitor land system dynamics (Kuemmerle et al 2013), has drawn recent attention. An added value of the use of radar, is understanding successional dynamics following disturbances. Since the magnitude of backscatter change is not linearly related to a change in biomass, a suitable interpretation of the relation is required. For example, clearances for large-area pastures and gold-mining were most common in lands with low backscatter before disturbance, and showed a large-magnitude drop in backscatter during disturbance, suggesting that large-scale deforestation was likely in areas with initially low biomass. Such areas may already be in use and in transition to a deforested state, and/or naturally have low biomass. We also found that the size of clearances and the land use type influences backscatter recovery. Large clearances for pastures recover fast, suggesting that the reversibility of conversion and accumulation of biomass post-disturbance may be an important sink for atmospheric carbon. In comparison, areas cleared for gold-mining recover significantly slower. Since post-disturbance increases in backscatter and biomass may be affected by a large number of factors (e.g. soil conditions, climatic variability and management practices), successional forest dynamics require much further analyses than those presented here: radar backscatter only provides a first-step in monitoring these as an integral part of deforestation and degradation processes.