Quantifying long-term changes in carbon stocks and forest structure from Amazon forest degradation

Despite sustained declines in Amazon deforestation, forest degradation from logging and fire continues to threaten carbon stocks, habitat, and biodiversity in frontier forests along the Amazon arc of deforestation. Limited data on the magnitude of carbon losses and rates of carbon recovery following forest degradation have hindered carbon accounting efforts and contributed to incomplete national reporting to reduce emissions from deforestation and forest degradation (REDD+). We combined annual time series of Landsat imagery and high-density airborne lidar data to characterize the variability, magnitude, and persistence of Amazon forest degradation impacts on aboveground carbon density (ACD) and canopy structure. On average, degraded forests contained 45.1% of the carbon stocks in intact forests, and differences persisted even after 15 years of regrowth. In comparison to logging, understory fires resulted in the largest and longest-lasting differences in ACD. Heterogeneity in burned forest structure varied by fire severity and frequency. Forests with a history of one, two, and three or more fires retained only 54.4%, 25.2%, and 7.6% of intact ACD, respectively, when measured after a year of regrowth. Unlike the additive impact of successive fires, selective logging before burning did not explain additional variability in modeled ACD loss and recovery of burned forests. Airborne lidar also provides quantitative measures of habitat structure that can aid the estimation of co-benefits of avoided degradation. Notably, forest carbon stocks recovered faster than attributes of canopy structure that are critical for biodiversity in tropical forests, including the abundance of tall trees. We provide the first comprehensive look-up table of emissions factors for specific degradation pathways at standard reporting intervals in the Amazon. Estimated carbon loss and recovery trajectories provide an important foundation for assessing the long-term contributions from forest degradation to regional carbon cycling and advance our understanding of the current state of frontier forests.


Introduction
Changes in Amazon forest carbon stocks are a significant source of greenhouse gas emissions from human activity (van der Werf et al 2009, Pan et al 2011, Aguiar et al 2016. Understanding the long-term response of Amazon forests to land use and climate is essential for balancing the global carbon budget and improving climate projections (e.g. Gatti et al 2014, Friedlingstein et al 2014. Although annual deforestation rates in the Brazilian Amazon have declined by 80% since 2004 (Hansen et al 2014, INPE 2015, forest degradation from fire and logging remains a threat to forest carbon stocks across the Amazon arc of deforestation (Morton et al 2013). The magnitude of carbon losses from forest degradation is large (Longo et al 2016), but the long-term consequences of fire and logging on forest structure and composition remain uncertain (Andrade et al 2017).
Decades of Amazon frontier expansion have left a mosaic of degraded forests along the Amazon arc of deforestation (Asner et al 2005, Morton et al 2013. Nearly 3% of southern Amazonia burned between 1999-2010, and the persistence of burned frontier forests (Morton et al 2013) underscores the importance of considering fire separately from deforestation for complete forest carbon accounting. Selective logging is also widespread across the leading edge of frontier expansion. In 2009 alone, 14.2 million m 3 of round wood was extracted from the largest logging centers in the Brazilian Legal Amazon (Pereira et al 2010). Canopy damage in logged forests can increase vulnerability to additional disturbances, including fire Vieira 1989, Holdsworth andUhl 1997), but the feedbacks and synergies among disturbance agents, as well as the long-term impacts of degradation, are still largely unresolved.
The scarcity of large-scale, long-term studies on fire and logging impacts has undermined efforts to quantify emissions from Amazon forest degradation for global carbon accounting (Le Queré et al 2016) and climate mitigation efforts (Andrade et al 2017). Reducing land-use emissions is one cost-effective climate mitigation pathway (e.g. Canadell andRaupach 2008, Griscom et al 2017), including efforts to reduce emissions from deforestation and forest degradation (REDD+) under the United Nations Framework Convention on Climate Change. To be eligible for REDD+ performance-based payments, countries must be able to monitor, report, and verify (MRV) reductions in carbon emissions from degradation or deforestation. However, because of large uncertainties regarding net carbon emissions from fire and logging, degradation has remained poorly integrated within the REDD+ accounting framework (Mertz et al 2012, Goetz et al 2015 and excluded from national reporting (e.g. Brazil 2014).
The challenge to quantify degradation emissions stems from the heterogeneity and time-dependence of degradation impacts (Longo et al 2016, Andrade et al 2017. The variability in degradation impacts may result from regional differences in underlying biomass distributions (Avitabile et al 2016, Longo et al 2016), forest resilience to fire (Brando et al 2012, Flores et al 2017, and land use (Aragão and Shimabukuro 2010). Discrepancies in emissions estimates also stem from methodological differences among studies. Field-based studies provide valuable context for understanding the long-term impacts of degradation (e.g. Berenguer et al 2014), but forest inventory measurements typically have limited spatial and temporal coverage due to cost constraints. By contrast, experimental studies control for much of the variability in degradation history but may be limited in their capacity to simulate the diversity of degradation impacts (e.g. Brando et al 2014).
Consequently, existing estimates for committed carbon emissions from Amazon understory fires vary by an order of magnitude, ranging from ∼20 Mg C ha −1 (Brando et al 2014) to 263 Mg C ha −1 (Alencar et al 2006). Airborne lidar provides the spatially extensive and structurally detailed information on forest structure and aboveground carbon stocks needed to reconcile previous estimates of degradation emissions and quantify co-benefits of avoided degradation (Goetz et al 2015, Longo et al 2016, Sato et al 2016. Here, we used a purposeful sample of high-density airborne lidar to capture a broad range of degraded and intact forest conditions in the southern Brazilian Amazon. For each forest stand, we combined degradation history information from annual time series of Landsat data with airborne lidar data to characterize canopy structure and estimate aboveground carbon density (ACD) using a lidar-biomass model specifically developed for frontier forests in the Brazilian Amazon (Longo et al 2016). Our large-area lidar coverage and sampling chronosequence addressed two questions: (1) What are the trajectories of loss and recovery of forest carbon stocks and habitat structure following fire and logging in frontier Amazon forests? (2) How do degradation type, frequency, and severity contribute to variability in degraded forest carbon stocks and habitat structure over time? Our study directly targets a lingering data gap for REDD+ (Andrade et al 2017) by quantifying the rates of ACD recovery over 1-to 15-year time horizons following a broad range of degradation pathways, including sequential impacts of logging and burning. These time-varying emissions estimates, or emissions factors, can be combined with activity data on the extent of forest degradation to establish REDD+ baselines; confirm the relative contributions from fire, logging, and regeneration to regional net forest carbon emissions; and estimate the consequences to mitigation targets if degradation remains omitted from greenhouse gas accounting. Airborne lidar also provides detailed, quantitative information on habitat structure that may support an improved understanding of the biodiversity co-benefits of reducing forest degradation-an integral, but poorly formalized component of REDD+ MRV.

Study area
The study area covers approximately 20 000 km 2 at the southern extent of closed-canopy Amazon forests in the Brazilian state of Mato Grosso (figure 1). Mean annual precipitation (1895 mm) and temperature (25 • C) support tropical forests and a diversity of land uses (Souza et al 2013). A four-month dry season (figure S1 available at stacks.iop.org/ERL/13/ 065013/mmedia) and periodic drought events (Chen et al 2011) contribute to the extent, duration, and severity of understory forest fires in the study region

Data and analysis
We combined Landsat time series and airborne lidar data to quantify variability in forest structure and ACD across gradients of degradation type, frequency, severity, and timing. Degradation history for areas with lidar coverage was characterized using a two-tiered classification approach. First, the annual occurrence of logging, understory fires, and deforestation was mapped based on spatial, spectral, and temporal information derived from annual time series of cloud-free Landsat mosaics for the early dry season months (June-August) of 1984-2016 (figure S2; text S1). Understory fires and deforestation events were identified based on multi-year patterns of damage and recovery in Landsat Normalized Difference Vegetation Index (NDVI) ( Mutually exclusive classification rules for the magnitude, duration, size, and shape of deforestation and degradation events avoided double counting errors common with the integration of independent products (figure S2; text S1) (Morton et al 2011, Bustamante et al 2016. Second, forest stands of uniform degradation history were manually delineated within the extent of lidar coverage and visually validated to confirm the extent and timing of degradation events. Logging roads visible in multiple years of Landsat data were excluded from logged forest stands to control for the impact of logging infrastructure on estimated carbon stocks and recovery trajectories. Airborne lidar data were used to estimate ACD in intact and degraded forest types stratified by degradation history. High-density airborne lidar data (minimum of 14 returns per m 2 ) were collected as part of the Sustainable Landscapes Brazil project across a range of intact and degraded forests in a space-for-time substitution sampling design (table S1, data available from: www.paisagenslidar.cnptia. embrapa.br/webgis/). Based on the classification approach described above, the 2891.25 ha of lidar coverage were stratified into 58 forest stands (4.50-498.50 ha; table S2).
A lidar-biomass model based on mean top of canopy height (TCH, m) (Longo et al 2016) was used to estimate ACD (kg C m −2 ) in forest stands at 0.25 ha resolution: where the parenthetical values are the standard errors of the parameters. Equation (1) assumes a biomassto-carbon conversion factor of 0.5, following Baccini et al (2012). We selected the TCH model because of its simplicity, sensitivity to the lower range of the ACD distribution, and accurate representation of ACD in burned forests (Longo et al 2016). Equation (1) was developed using inventory and lidar data from intact and degraded Amazon forests. Here, we applied the model to a new set of lidar data sampled from the same regional context in which the Longo et al (2016) model was calibrated; about 8% of the lidar data set overlapped with the data used in model development.
Pixel-based uncertainty associated with modeled ACD was calculated from three sources of statistical uncertainty following the methods described in Longo et al (2016). A Monte Carlo approach with 10 000 iterations was used to propagate the pixelbased uncertainty to the stand level by adjusting each biomass pixel with randomly distributed noise proportionate to its uncertainty before aggregating data at the stand level. The stand-level standard error was derived from the standard deviation of the simulated standlevel means.
Given the importance of canopy structure for wildlife habitat in tropical forests (Bergen et al 2009), we also calculated two lidar-based measures of habitat structure. First, residual canopy cover was calculated using 1 m resolution lidar canopy height models (CHMs) as the proportion of the forest stand greater than or equal to the mean canopy height in intact forests (21 m). Second, clusters of one or more canopy trees (≥21 m) were identified using the 1 m CHMs with a maximum search radius of 10 m using a 3 × 3 pixel moving window (Silva et al 2015). These metrics provided complementary information on changes in forest structure from degradation and recovery processes to assess the drivers of ACD variability and the timevarying recovery of both carbon and habitat structure in degraded forests.
We used multiple linear regression to model the loss and recovery trajectories of ACD and canopy structure based on the chronosequence of lidar samples. Four least squares models were fit using the lm function in R version 3.3.0 (www.R-project.org). Model 1 estimated median ACD in degraded forest stands based on degradation type (burned or logged-only), timing (years since last degradation event), and fire frequency. Median ACD was selected as the measure of central tendency for each stand because of the skewed ACD distributions in degraded forests. Model 2 further stratified once-burned forests by fire severity, visible as rings of high-and low-severity canopy damage, based on the relative difference between the pre-fire and post-fire Landsat dry-season NDVI (RdNDVI). A fixed threshold of mean minus the standard deviation of RdNDVI was only used to stratify low and high-severity fire damages in once-burned stands because the spatial variability of fire damages was not well preserved following recurrent fire events. Models 3 and 4 used degradation type, timing, and frequency to predict residual canopy cover and density of canopy tree clusters, respectively. In all four models, the variable for time since last degradation event was log-transformed to satisfy assumptions of normality and homoscedasticity (Vargas et al 2008, Becknell et al 2012. Additionally, to isolate the effect of forest recovery from the long-term impacts of logging infrastructure, logged forest stands were adjusted to exclude secondary roads and log landing decks. Interactions between degradation history (type, frequency, severity) and degradation timing were evaluated for significance and model performance in all four models. Lastly, differences across degradation strata were evaluated using pairwise Wilcoxon tests to accommodate the diversity of non-normal data distributions.
Consistent with recommendations from the Intergovernmental Panel on Climate Change (Penman et al 2003), an additional Monte Carlo procedure was used to propagate the effect of ACD uncertainty on model parameters and predictions by performing 10 000 realizations of the model fit on adjusted stand-level medians with normally distributed noise proportional to the stand-level standard error, or the standard deviation of the stand-level Monte Carlo aggregations.

Results
Degradation type, frequency, timing, and severity contributed to ACD variability in frontier forests. Lidar-based estimates of ACD in 58 Amazon forest stands varied by nearly two orders of magnitude between the most heavily degraded forest stand (median: 4.5 Mg C ha −1 ), a stand that had been logged and burned three times, and the most carbon-dense intact forest stand (median: 114.3 Mg C ha −1 ; table S2). At the pixel scale, median carbon density in degraded forests (51.2 Mg C ha −1 ) was less than half of ACD in intact forests (113.5 Mg C ha −1 ). Degraded ACD was also more heterogeneous than intact ACD (coefficient of variation: 68.4% and 16.7% for degraded (2638.00 ha) and intact forest pixels (253.25 ha), respectively).
The variability in ACD following degradation could not be constrained by degradation type alone. ACD in pixels with a history of fire (median: 20.4 Mg C ha −1 ; Fire frequency ACD normalized to intact forest (%) Figure 2. Forests affected by multiple fires had the largest differences in aboveground carbon density (ACD) compared to median ACD in intact forests (red line). Additionally, ACD distributions between burned and logged-and-burned forests were similar for once-burned forests. The violin plots summarize ACD distributions as a function of fire frequency and degradation class. The median and interquartile range of each group are indicated with the black circle and line. The tails of the violins are trimmed to the range of data, and all violins have the same area prior to trimming the tails. Distinct letters indicate significant differences among distributions from a pairwise Wilcoxon test with a Holm correction procedure to adjust for multiple testing. See figure S4(a) for a comparison across burn frequency groups that also accounts for stand age.
1605.75 ha) was significantly lower (p < 0.05) than ACD in logged-only pixels (77.8 Mg C ha −1 ; 1032.25 ha); however, ACD varied broadly within both degradation classes. At the stand level, there was considerable overlap between the ranges of median ACD in burned forests (4.5-95.2 Mg C ha −1 ) and logged-only forests (39.0-117.3 Mg C ha −1 , table S2).
Degradation timing was a critical factor for further differentiating ACD between and within logged and burned forest classes (figure S3; table 1). Within two years of recovery, median ACD in burned pixels was 9.5 Mg C ha −1 , compared to 68.4 Mg C ha −1 in logged-only pixels. Following 10 to 15 years of recovery, neither class recovered its estimated pre-disturbance ACD, and median ACD in burned pixels remained considerably lower than in logged pixels (difference: 17.5 Mg C ha −1 ).
Fire frequency governed both the magnitude and the spatial pattern of residual forest carbon stocks (figures 2 and 3; table 1). Repeated burning resulted in a non-linear decline in ACD, irrespective of logging history, with lowest ACD in forests subjected to three or more fires (figure 2). Forests affected by a single fire (n = 10) retained 67.0 Mg C ha −1 (interquartile range [IQR] ± 26.4 Mg C ha −1 ). Twice-burned forests (n = 5) contained less than half the carbon stocks in once-burned forests (31.6 ± 21.1 Mg C ha −1 ). Forests burned three to five times (n = 13) retained few trees from the pre-fire forest stand; ACD was only onesixth of that of once-burned forests (10.3 Mg C ha −1 ), with the narrowest IQR of all burn frequencies (±10.5 Mg C ha −1 ). Importantly, the observed decrease in IQR with increasing fire frequency indicated a reduction in structural complexity from repeated burning (figures 2 and 3).
Unlike the impact of successive fires, there was no significant long-term impact on ACD recovery attributable to prior logging after controlling for fire frequency ( figure S4). Because the distinction between burned and logged-and-burned forests was not a statistically significant predictor of degraded forest ACD, nor did it improve model fit, logged-and-burned and burned forest stands were combined to model post-fire recovery of ACD.
Fire frequency and the time since the last degradation event explained the greatest variability in degraded ACD recovery (Model 1; adjusted R 2 = 0.89; F-statistic = 106.5 figure 4(a); table S3). The immediate reduction in ACD differed significantly for each degradation pathway (regression intercept; table S3).
In the year following degradation, the modeled ACD for forests that had been logged, once-burned, twiceburned, and subjected to three or more burns was 62.3, 52.0, 19.4, and 11.0 Mg C ha −1 , respectively. However, the rate of ACD recovery was similar for all classes, as interaction effects between fire frequency and time since degradation event were not statistically significant (table S3). Given these initial differences and the slow recovery in degraded forest ACD, the legacy of forest degradation was still evident 15 years following fire and logging (table 2).
Initial fire severity was a statistically significant predictor of ACD recovery in once-burned forests (Model 2; Adjusted R 2 = 0.88; F-statistic = 87.87; figure 4(b); tables 2 and S3). In the year following fire, estimated high-and low-severity damages differed by 16% of intact ACD (table 2). Modeled differences in ACD resulting from initial fire severity were preserved through time, with once-burned forests recovering between 57.6% and 73.9% of intact ACD after 15 years of recovery, depending on initial fire severity (figure 4(b); table 2). Covariation of ACD with Landsat and lidar metrics of canopy density in burned forests provided additional insights into the contribution of fire Table 1. Forest degradation from logging and fire alters ACD and stand structure relative to neighboring intact forests. Lidar-based estimates of the fraction of original canopy cover, number of canopy tree clusters, and the distribution of ACD in degraded forests. Degraded forests were partitioned along three axes of variability-degradation type, frequency, and timing. The lower, middle (median) and upper quartile of aboveground biomass density (Mg C ha −1 ) are shown as ACD 25 , ACD 50 , and ACD 75 , respectively. -30.9 -0 ACD NDVI Figure 3. Ring patterns in burned forests indicate diurnal differences in fire line intensity, and increasing fire frequency results in a progressive loss of forest biomass and structural diversity. Lidar-based estimates of aboveground carbon density (ACD, Mg C ha −1 ) at 0.25 hectare resolution for 5000 × 200 m transects are overlaid on post-fire Landsat NDVI for once-burned (a) twice-burned (b) and thrice-burned (c) forest stands. See tables S1 and S2 for additional profile information associated with each stand (stand IDs from left to right: 26, 13, and 8). Table 2. Estimates based on the multiple linear regression models of aboveground carbon density predicted at four standard reporting periods following the most common logging and fire pathways. For each degradation class, modeled ACD and 95% confidence interval (in parentheses) are shown as the percentage of the intact forest reference (113.5 Mg C ha −1 ). The confidence interval was calculated based on the mean of 10 000 confidence intervals generated from the Monte Carlo linear regressions, which were iteratively fit to the stand-level biomass estimates adjusted with noise proportionate to the stand-level standard errors. Model predictions for low-and high-severity fires are derived from model 2; all other predictions presented here are derived from model 1 (see   ; table  S4). Understory fires resulted in the largest reduction of canopy tree clusters, particularly following recurrent fires. Logged forests retained more than twice as many canopy tree clusters (46.5%) as once-burned forests (20.0%) when measured within 1-2 years of the degradation event. Forests burned three or more times retained only 4.7% the number of canopy tree clusters found in intact forests. After 14-15 years of regrowth, once-burned forests recovered only 80% of the canopy tree clusters present in logged forests. Further, these impacts to forest structure may persist even after ACD in degraded forests returns to predegradation levels. For example, after 14-15 years of regrowth, once-burned forests recovered a larger fraction of intact-forest ACD (80.2%) than canopy tree clusters (58.2%).

Discussion
Amazon forest degradation from logging and fire has a lasting impact on forest carbon stocks and canopy structure. The slow recovery of degraded forests under-scores the need to address drivers of degradation to ensure the retention of carbon stocks and preserve complex canopy structure in frontier Amazon forests. Using a large sample of intact and degraded forests, we provide the first comprehensive look-up table of degradation emissions factors for Amazon forests to guide the incorporation of forest degradation within REDD+ MRV (tables 1 and 2). Our findings illustrate the persistence of degradation impacts beyond the time scales for REDD+ MRV and even REDD+ baselines (typically 10 years), providing the foundation for further investigations into the relative contributions from fire and logging to regional land-use emissions. ACD in degraded forests varied by two orders of magnitude across the study area (table S2), providing clear support for the creation of multiple classes of forest degradation within REDD+ or other carbon accounting frameworks based on degradation frequency, severity, and timing. Overall, understory fires led to larger and more persistent changes in ACD and forest structure than logging, consistent with previous findings from Longo et al (2016). Our results further demonstrate how fire severity and fire frequency contribute to non-linear declines in ACD and homogenization of degraded forest structure (figure 2, tables 1 and 2). Collectively, these results address key data gaps that have hindered MRV of Amazon forest degradation.
Lidar-based estimates of carbon losses from fire were much larger than previous reports from experimental studies and forest inventories. For example, the reduction in ACD one year following a single burn Years since last degradation ACD Mg C ha −1 Figure 4. Patterns of aboveground biomass recovery following forest degradation highlight the magnitude and duration of ACD accumulation following fire. (a) Relationship between ACD and stand age for forests that were logged, burned once, burned twice, and burned three or more times. (b) Initial fire severity in once-burned forests further explains the heterogeneity in residual carbon stocks. Points correspond to estimated stand-level medians, error bars correspond to stand-level standard errors derived from 10 000 Monte Carlo stand-level aggregations, and the shaded bands represent the mean 95% confidence interval from 10 000 Monte Carlo simulations of the model fit. Model details are presented in table S3.
in this study (54.2%) was approximately three times larger than from experimental fires in the southeastern Amazon (Brando et al 2014). This discrepancy may reflect the improved capacity to characterize the heterogeneity of wildfire damages using airborne lidar or the difficulty for prescribed fires in experimental studies to replicate the emergent properties of wildfires, such as fire front intensity. Field studies have also reported smaller relative losses in ACD following fire (13.7%; Berenguer et al 2014). These differences may reflect the confounding influence of different age classes and burn frequencies, the challenges of capturing the length scales of spatial variability (see figure 3) using typical inventory plots (0.25-1.0 ha), or regional variability in fire intensity from climatic and forest-type specific responses to fire (e.g. Flores et al 2017). These broad discrepancies reinforce the need for large-scale studies of additional frontier landscapes to support emissions mitigation programs, including REDD+ MRV.
Reducing the incidence and frequency of understory forest fires would preserve both carbon stocks and habitat structure in frontier landscapes. The marginal carbon cost of recurrent fire events in this study suggests that avoiding just one additional fire in a previously burned forest would retain carbon stocks equivalent to one-third of the intact reference ACD. Notably, not all degradation sequences have the same cumulative impact. We contrast the non-linear impact of recurrent burns with the effect of selective logging before fire. In the case of recurrent burns, each fire leads to a greater proportional loss. However, logging before fire did not amplify the long-term carbon losses from fire, after accounting for fire frequency; nor was logging a significant predictor of carbon recovery, regardless of fire history. These findings suggest that the distribution of fine litter (e.g. Balch et al 2008) may be a more important determinant of fire damage than large woody debris or canopy openings from logging.
The slow recovery of degraded ACD suggests that the continued omission of degradation from carbon accounting may result in substantial underreporting of forest carbon emissions. Relative to baseline periods, the frequency and severity of Amazon droughts (Boisier et al 2015, Duffy et al 2015 are projected to increase degradation risk in coming decades (Nobre et al 2016, Le Page et al 2017. The look-up table of proportional losses between degraded and intact forests developed in this study may facilitate the integration of carbon losses from fire and logging into REDD+ monitoring and reporting protocols. Further, accounting for carbon emissions from forest degradation may also reduce uncertainties in the Amazon carbon budget. Previous studies have either excluded a post-disturbance recovery term (Aragão et al 2014) or have combined secondary and degraded forests (Houghton et al 2000, Pan et al 2011, despite the diversity of loss and recovery pathways among degraded and secondary forest types (Poorter et al 2016).
Parallel ACD recovery curves in years 1-5 following logging and fire may reflect common site constraints, distinct mechanisms of forest growth, and model calibration. For example, different mechanisms of vegetation recovery and canopy closure may generate similar changes in estimated ACD, such as small gains in mean canopy height in logged forests and fast height growth of shorter resprouting or surviving trees in burned forests. Additionally, given that logging intensity is the single best predictor of ACD recovery time (Rutishauser et al 2015), evidence for greater extracted wood volume of low-value species in frontier forests (Richardson and Peres 2016) than in interior forests and experimental logging sites may explain differences with previous estimates of ACD recovery in logged forests (e.g. Chambers et al 2004, Putz et al 2012, Andrade et al 2017. Further, moisture availability is a critical constraint on regeneration rates (Poorter et al 2016, Wagner et al 2016; moisture stress from the seasonality of the study site may limit recovery rates in both logged and burned forests. Additional observations in repeatedly burned forests are needed to constrain long-term estimates of recovery patterns (>5 years) in the more heavily degraded sites.
Airborne lidar captures details about 3D forest structure needed to quantify aboveground carbon stocks and advance quantitative reporting on biodiversity safeguards and other co-benefits of REDD+. Individual tree and plot-level data from airborne lidar provide insights into the mechanisms driving biomass variability and habitat impacts from forest degradation. The residual density of large canopy trees, which can be directly quantified using high-density airborne lidar, is an important driver of ACD variability in degraded forests (Slik et al 2013), and closely corresponds to the spatial patterns of fire-induced canopy mortality ( figure S7). In addition to ACD, the loss of canopy trees may also alter the forest micrometeorology, aerodynamic roughness, and successional success of grasses and lianas (Ray et al 2005, Silvério et al 2013. These changes, in turn, can increase vulnerability to windthrow and repeated fires, especially during drought years (e.g. Balch et al 2015). Canopy trees also serve as biodiversity refugia; the slower recovery of canopy tree clusters than carbon stocks in this study may suggest a more persistent impact of degradation on biodiversity than biomass in the first decades following logging or fire, consistent with findings from Martin et al (2013). Characterizing the time-integrated effects of avoided degradation on forest structure is clearly an important step for policies and management that aim to promote the retention of both biomass and biodiversity. Measurement and monitoring capabilities to support REDD+ commitments to safeguard biodiversity and promote other co-benefits are not yet operational (Goetz et al 2015). This work highlights the potential of airborne lidar to advance REDD+ MRV for both carbon and non-carbon objectives.
Our findings provide a detailed characterization of the carbon and habitat changes following Amazon forest degradation, but additional measurements are needed to assess regional variability in degradation impacts. Additional lidar samples across gradients in land use, forest type, and climate may identify important differences in degradation impacts and ACD recovery. For example, previous work suggests that transitional forests along the southern extent of the Amazon may be more resilient to mortality from a single, low-severity fire during average weather conditions (Brando et al 2012) than interior forests. By contrast, forests in Central Amazon floodplains have exposed roots during dry periods, thin bark, and lack the ability to resprout, rendering them more vulnerable to fire-induced dieback (Flores et al 2017). Additionally, multi-temporal observations are needed to unequivocally attribute ACD losses to degradation, characterize delayed mortality, and investigate the potential for arrested succession (Barlow et al 2003). Multi-temporal studies may also help constrain interannual variability in fire damages (Brando et al 2014), consistent with the ∼15% difference in ACD observed in this study between low and high-severity damages within a single fire. Complementary field measurements may help characterize key aspects of degraded forest structure that are not well captured by airborne lidar, such as the species distribution of regeneration from seeds or sprouts and the selective impact of degradation on mean wood density (Bunker et al 2005, Longo et al 2016. Lastly, the strong correspondence between changes in Landsat surface reflectance and lidar-derived estimates of forest structure and ACD in burned forests may support regional estimates of carbon losses from understory fires using Landsat or similar moderate resolution imagery.

Conclusion
Forest degradation is ubiquitous in frontier Amazon forests, and damages from logging and fire were larger and longer lasting than previously reported for our southern Amazon study region. Combining the lookup table of emissions estimates from this study with activity data from satellite monitoring programs may allow for regional estimates of combined emissions from deforestation and forest degradation for REDD+. Understory fires-particularly, repeated burns-pose the greatest risk to forest carbon stocks and canopy structure along the Amazon arc of deforestation. Thus, avoiding additional fires in frontier landscapes may have an outsized benefit for carbon retention and habitat. Routine monitoring of frontier forests with airborne lidar may provide additional insights regarding the direct impacts of forest degradation on both carbon stocks and forest structure, including potential interannual variability from climate controls on fire severity or market influences on logging removals. Our approach to disentangle the complex legacy of degradation by combining forest inventory, airborne lidar, and Landsat time series offers a blueprint to generate degradation emissions factors in other geographies and regional circumstances.