Impact of deforestation and climate on the Amazon Basin’s above-ground biomass during 1993–2012

Since the 1960s, large-scale deforestation in the Amazon Basin has contributed to rising global CO2 concentrations and to climate change. Recent advances in satellite observations enable estimates of gross losses of above-ground biomass (AGB) stocks due to deforestation. However, because of simultaneous regrowth, the net contribution of deforestation emissions to rising atmospheric CO2 concentrations is poorly quantified. Climate change may also reduce the potential for forest regeneration in previously disturbed regions. Here, we address these points of uncertainty with a machine-learning approach that combines satellite observations of AGB with climate data across the Amazon Basin to reconstruct annual maps of potential AGB during 1993–2012, the above-ground C storage potential of the undisturbed landscape. We derive a 2.2 Pg C loss of AGB over the study period, and, for the regions where these losses occur, we estimate a 0.7 Pg C reduction in potential AGB. Thus, climate change has led to a decline of ~1/3 in the capacity of these disturbed forests to recover and recapture the C lost in disturbances during 1993–2012. Our approach further shows that annual variations in land use change mask the natural relationship between the El Niño/Southern Oscillation and AGB stocks in disturbed regions.

may lead to a decrease (increase) in AGB obs stocks, resulting in positive (negative) trend in AGB def . Meanwhile, the CO 2 -fertilization effect may lead to a greater potential for forest regeneration (i.e. greater AGB pot ) as recent findings indicate it is the main driver of a global greening of the land surface 14 . However, locally changing climate conditions may lead to a reduction of the resilience of tropical forests and a transition toward less densely vegetated savannah landscapes 15 . There is a projected risk of Amazon die-back 7 due to climate change, albeit with large uncertainty on its occurrence and severity 16 . It would reduce the potential for biomass recovery associated with reforestation by the end of the 21 st century. Therefore, it is important to estimate the resilience of AGB pot to climate change to design efficient climate mitigation strategies based on reforestation.
In this study, we build on a previous approach 8 (see Methods) to address the evolution of AGB pot , and hence AGB def , using a new dataset 17 that provides annual estimates of AGB obs from 1993 to 2012 at a 0.25° spatial resolution. By doing so, we aim to answer the following questions: • How did AGB def evolve in disturbed regions of the Amazon Basin over these two decades?
• Can we apportion this evolution to climate conditions affecting AGB pot versus human activities reducing AGB obs ? • Would reforestation-based mitigation strategies be resilient to climate change in previously cleared regions of the Amazon Basin?

Results
We estimate a change in AGB obs from 26.3 Pg C (with a 4.1 Pg C confidence range) in 1993 to 24.1 Pg C (with a 3.9 Pg C confidence range) in 2012, or a 2.2 Pg C (with a 0.2 Pg C confidence range) loss in regions of the Amazon basin which are not IFL. Using the machine-learning approach we derive a reduction of AGB pot from 32.1 Pg C (with a 4.0 Pg C confidence range) in 1993 to 31.4 (with a 3.9 Pg C confidence range) in 2012 in the same regions. Comparing the evolution of AGB obs and AGB pot results in a human-driven increase in AGB def from 18.0% (AGB def /AGB pot ) in 1993 (with a 2.3% confidence range) to 23.3% in 2012 (with a 2.7% confidence range). Overall, ~1.5 Pg C of the ~7.3 Pg C mean AGB def in 2012 was generated by combined anthropogenic activities and climate patterns since 1993 (Table 1). The evolution of AGB def is strongly linear during 1993-2005 (r = 0.99; p ≪ 0.001) before plateauing from 2005 onwards with no significant trend (Fig. 1). The stabilisation of AGB def after 2005 is associated to a reduction of AGB obs loss rate from 0.17 Pg C y −1 (with a 6% relative uncertainty) to 0.04 Pg C y −1 (with a 14% relative uncertainty) before and after 2005 respectively (Fig. 2). It corresponds to a reduction in deforestation rates over the Brazilian Amazon seen in data from INPE ( Figure S1 in the Supplementary Information; r = 0.97; p ≪ 0.001) while the smooth decreases of AGB pot throughout the study period indicates a long-term negative impact of climate on the regeneration potential of disturbed regions (Fig. 2). The increase in AGB def is heterogeneously distributed across disturbed areas of the basin (Fig. 3). While the spatial distributions of AGB def are significantly correlated (r = 0.89; p ≪ 0.001) in 1993 (Fig. 3a) and 2012 (Fig. 3b),   Fig. 3c). We note a reduction in AGB def , i.e. a recovery of AGB obs stocks toward AGB pot , in the south-eastern edge of the basin, and to a lesser extent in northern Brazil. This recovery indicates that non-primary vegetation, mostly rangeland in these regions, may have built up biomass stocks from 1993 to 2012. Over the period 1993-2012, local increases in AGB def can be explained by the erosion of primary land (Fig. 4). Conversely, local recovery of stocks associated to decreases in AGB def corresponds to regions where the fraction of primary land was already low in 1993. This pattern indicates a recovery of AGB stocks in other land cover types, principally rangelands ( Figure S2). Despite this apparent recovery of AGB stocks, the deficits in these regions were still > 50 Mg C ha −1 in 2012.
Our estimates indicate a significant negative correlation between inter-annual variations of the El Niño/ Southern Oscillation (ENSO), represented by a winter composite of the Multivariate ENSO Index (MEI w , see methods) and detrended ΔAGB pot integrated over previously disturbed regions ( Figure S3 in the Supplementary Information; r = −0.57; p ≈ 0.01). This relationship indicates that negative (La Niña) phases of ENSO would drive positive anomalies in ΔAGB pot , i.e. a stronger sink, while positive (El Niño) phases of ENSO are associated with negative anomalies in ΔAGB pot , a weaker sink. However, past and current human activities mean that this significant relationship between ENSO and the sink strength disappears when comparing with de-trended ΔAGB obs (r = −0.38, p > 0.10). We conclude that, through clearing and subsequent regrowth, human activities have become the main driver of inter-annual variability of the land-based sink, dominating natural climate drivers, in disturbed regions of the Amazon.

Discussion
The annual biomass maps have allowed resolution of AGB changes across the Amazon Basin, indicating areas of heavy losses, but also some areas of AGB gain (Fig. 3). By mapping the potential biomass, we show the evolution of the basin's capacity to store C, a baseline without human impacts. Because AGB pot is determined from annual AGB obs data in IFL, the annual variation in AGB pot indicates the effect of climate on the storage capacity of the intact forest. We show that this potential has declined over 1993-2012 ( Fig. 2) similarly to AGB stocks in IFL ( Figure S4 in the Supplementary Information), due to climate and in spite of rising atmospheric CO 2 concentrations (Table 1). Indeed, the evolution of AGB stocks in IFL is significantly correlated with the vegetation water stress estimated by GLEAM 18 20,21 . Overall, these results indicate that drying conditions have degraded the capacity of the disturbed regions to regain their lost biomass which is line with the projected risk of climate driven Amazon biomass loss 7 . This climate-driven reduction in the capacity for regeneration also corroborates with risks for tropical forests to be replaced by savannahs if drier conditions dominates 15 .
Our results are first-order estimates and we are aware that hard-to-quantify and potentially large uncertainties may arise from ground-level measurements 22 , the way they are used in combination with remote-sensing data to derive large-scale biomass maps 23 , and the identification of forest cover 24 and intact forest landscapes 13 . Therefore, we have validated the robustness of our machine-learning approach in several ways. First, it simulates annual AGB obs with <0.1% bias integrated over out-of-sample IFL regions ( Figure S5a in the Supplementary Information). We note a tendency to overestimate AGB in less densely vegetated regions ( Figure S5b,c in the Supplementary Information) but the local mean relative bias is <1.2%. Second, pixel to country-scale estimates of the evolution of AGB def through time are in agreement with independent datasets of deforestation ( Figure S1) and land cover change rates (Fig. 3). Finally, the ~7.3 Pg C AGB def estimated after 2005 is similar to the one reported previously 8 . Our highest confidence results indicate a ~0.08 Pg C y −1 increase in AGB def for the period 1993-2012. This net number is about half of recent estimates of gross C emissions from the Amazonian deforestation 25 . It is in agreement with the ~50% compensation of gross C emissions from tropical deforestation by regrowth 1 . Assuming that large-scale deforestation started in 1960 (ref. 26 ), the initial AGB def of ~5.8 Pg C in 1993 corresponds to a higher 0.18 Pg C y −1 net biomass loss prior to this date. The decrease in AGB def growth rate between 1993 and 2012, and especially after 2005 (Fig. 1), matches reports of a slowing down of Brazilian deforestation during 2005-2012 (refs [26][27][28] ) but is also a result of a decrease in AGB pot in disturbed regions of the Amazon Basin.
Furthermore, field studies 20,21 and airborne measurements 29 have shown that climate variability, and especially El Niño-induced droughts, have a large impact on the carbon balance of undisturbed areas of the Amazon Basin. These previous results are in agreement with the negative correlation between MEI w and ΔAGB pot ( Figure S3 in the Supplementary Information). Overall, human-induced clearing and recovery processes mask the natural response of ecosystems to climate in disturbed parts of the Amazon Basin. While this impact is intuitive, we are able to demonstrate it quantitatively with the AGB pot reconstructions. Finally, this result raises concerns on the viability of climate change mitigation strategies, as climate change is likely to challenge the resilience of forested landscapes.

Conclusion
We have recreated annual maps of potential AGB for the Amazon Basin, which allows the net impacts of global change on basin biomass to be determined. Compared to maps of historical biomass, these indicate an increase of ~1.5 Pg C in the biomass deficit (AGB def ) for 1993-2012. This basin-wide number is a net estimate of climate-induced variation of AGB pot and deforestation-induced erosion of AGB stocks, which are partly compensated by regrowth in some areas post-deforestation. Overall, our results indicate that land use change continues to erode the carbon storage of the Amazon basin while climate change is impairing its capacity to sequester carbon through natural processes of regrowth, raising concerns on the long-term resilience of land-based mitigation strategies.

Methods
Annual maps of AGB. We use annual Above Ground Biomass maps 17 (AGB obs ) for the period 1993 through 2012 based on the passive microwave observed vegetation optical depth (VOD, dimensionless) from a series of satellites. VOD is an indicator of the total water content in the aboveground vegetation, i.e. including both canopy and woody components [30][31][32] . This VOD dataset can qualitatively capture the long-term and inter-annual variations in vegetation water content over different land cover types [33][34][35][36][37] . Annual AGB obs maps were created by  establishing a relationship between VOD and a pan-tropical map 4 of AGB obs circa 2000. These annually resolved maps are comparable with previous independent estimates of AGB dynamics 1,5,6 . For more details about the methodology used to create AGB obs maps, please refer to Liu et al. (2015, ref. 17 ).
Creating potential AGB maps. To derive the evolution of the AGB deficit (AGB def ) we first created annually resolved maps of potential Above Ground Biomass (AGB pot ) in previously disturbed regions. AGB pot corresponds to AGB stocks there would exist under current climate if deforestation had not occurred in these regions. It can also be conceptualized as the current forest regeneration potential if regrowth was instantaneous. The method to create AGB pot maps was described in Exbrayat and Williams (2015; ref. 8 ) and is only briefly summarized hereafter.
First, we used a Random Forest machine-learning algorithm 38,39 to reproduce AGB obs as a function of climatology in identified Intact Forest Landscapes (IFL) which cover about 55% of the Amazon Basin. The Random Forest technique relies on multiple decision trees (here n = 1,000) to group data points as a function of driving data. Then, in each final node a multiple linear regression is trained to predict the target variable (here AGB obs ) as a function of explanatory data. Each individual decision tree is trained on a randomly selected subset of the data and the final prediction is the average of all trees. Here, we use the CRU CL2.0 climatology dataset 12 , re-gridded to a matching 0.25° resolution with the Climate Data Operators version 1.6.9, and latitude, a proxy of intra-annual photoperiod amplitude, as explanatory variables to predict AGB in IFL. The assumption is made that regions identified as 'intact' may be subject to small-scale indigenous management 40 or disturbances 41 that are negligible at the coarser 0.25° resolution used here 8 . Compared to our previous study we used an updated IFL dataset 13 that represents the extent of intact regions for the year 2013. It ensures that training regions have remained intact throughout the whole period covered by the AGB obs dataset (i.e. 1993-2012). In addition to these continuous drivers, we used a categorical variable to separate pixels corresponding to large-scale open water regions in the Global Lakes and Wetlands Database 42 . As VOD values are strongly influenced by the open water dynamics, the pixels with large-scale open water are identified and the VOD values over these pixels are assumed constant among different years 17 .
Once trained the algorithm can then be used to estimate annual, climate-driven, AGB pot in previously disturbed regions (i.e. outside IFL) regions. Although it has been identified as the major driver of the recent greening of the land surface 14 , CO 2 is not explicitly used in our approach because of the lack of availability of spatially-explicit data of atmospheric concentrations. However, we assume that the impact of increasing CO 2 on AGB stocks is intrinsically included in time series of AGB in IFL which also include the impact of changing climatic conditions. Using annual maps of AGB pot we can calculate an AGB deficit (AGB def = AGB pot − AGB obs ) and derive time series of its evolution from 1993 to 2012. As the temporal evolution of AGB pot is only driven by climate and atmospheric CO 2 concentrations, we assume that AGB def is representative of the net and cumulative impact of anthropogenic activities on biomass dynamics on AGB stocks. We perform the analyses using the mean AGB obs from Liu et al. (ref. 17 ) to derive AGB pot and AGB def . Furthermore, we evaluate the uncertainty in our approach by performing the analysis with the 5 th and 95 th percentiles of AGB obs data 17 to report the corresponding confidence ranges in AGB pot and AGB def . As a proof of concept, we first validate the method using ~50% of randomly selected pixels in IFL as training dataset and the remaining IFL pixels as target dataset to assess the robustness of the approach to recreate 20 years of AGB pot . Corresponding results are presented in Figure S5 of the supplement. We note a good agreement between reconstructions and data in IFL although there is a tendency for the machine-learning to overestimate AGB in less densely vegetated regions.
Validation of results. Our estimates of AGB pot cannot be directly validated against field data. However, we expect the temporal evolution of AGB def to be related to contemporary deforestation rates and land cover changes. Therefore, we compare time series of AGB pot from pixel to country-scale with independent datasets of Land Use and Land Cover Change (LULCC). First, we compare annual deforestation rates reported by INPE for the Brazilian part of the Amazon Basin with the corresponding trend in AGB def over the whole period 1993-2012. Second, we use spatially-explicit data from the Land-Use Harmonization project version 2 (LUH2v2h; data updated from ref. 43 ). LUH2v2h is a global driving dataset that provides annual land cover information for the period 850-2015 C.E. in the Land Use Model Intercomparison Project 44 (LUMIP) contribution to the upcoming sixth phase of the Coupled Model Intercomparison Project 45 (CMIP6). In LUH2v2h land covers are distributed between 12 classes (2 primary land classes, 2 secondary land classes, 5 cropland classes, 2 pasture and rangeland classes and 1 urban class) and the fraction they cover in each 0.25° pixel is reported annually.
Climate sensitivity. We compare the evolution of AGB obs in IFL with time series of the vegetation stress factor S from the GLEAM dataset v 3.1a (ref. 18 ). GLEAM is a data-assimilation system that uses satellite observations to constrain daily estimates of global terrestrial evaporation and root-zone soil moisture 46 . The factor S is an output of GLEAM and represents the ratio of actual evapotranspiration to potential evapotranspiration, an indicator of ecosystem's water stress. It is as a function of vegetation state and soil moisture availability and therefore takes long-term effects of precipitation conditions into account. We use the mean annual value of S across the IFL regions of the Amazon Basin, expressed as a z-score, to explain the evolution of AGB obs ( Figure S4).
We seek to further understand the impact of large-scale human disturbances by quantifying their impact on the response of ecosystems to climate variability. We focus on the El Niño/Southern Oscillation (ENSO), a main driver of global climate variability 47 . The state of ENSO, quantified through the calculations of an index, significantly correlates with the strength of the global land carbon sink 48 . Indeed, positive (negative) El Niño (La Niña) phases drive warmer and drier (cooler and wetter) conditions over large parts of the pan-tropical region, including the Amazon Basin, which explains spatial patterns of ecosystem carbon uptake 48 . Following previous studies 48, 49 we use a winter composite of the Multivariate ENSO Index 50,51 calculated between Dec/Jan and Mar/ Apr (referred as MEI w ). To quantify the impact of human disturbances on the response of the Amazon terrestrial carbon sink to ENSO, we study the correlation between MEI w and detrended anomalies of annual ΔAGB obs and ΔAGB pot stocks integrated over disturbed (i.e. non-IFL) regions of the Amazon Basin. We choose to rely on a global index rather than actual data of temperature and precipitation for the Amazon Basin because past deforestation may have altered these quantities in regions where land-atmosphere coupling is strong 52,53 . Data availability. The data generated during this study are available from the corresponding author on reasonable request.