Distinct drivers of extreme and non-extreme fires in the Brazilian Amazon

Distinct drivers of extreme and non-extreme fires in the Brazilian Amazon | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Distinct drivers of extreme and non-extreme fires in the Brazilian Amazon Wei Li, Zhixuan Guo, Philippe Ciais, Stephen Sitch, Guido van der Werf, and 8 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7604250/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted You are reading this latest preprint version Abstract The Brazilian Amazon suffers from frequent fires with varying sizes and intensities, but the causes for extreme and non-extreme fires remain unclear. Here, using multiple satellite-based observations and explainable machine learning models, we find distinct drivers of burned area variation for extreme (burned area > the 90th percentile) and non-extreme fires in the Brazilian Amazon between 1985 and 2020. The absolute land use fraction of pasture and forest are dominant drivers for non-extreme fires, while the extreme fires are more driven by the fire weather conditions and land use change from forest to pasture in the adjacent two years. In future climate and land use change scenarios, our predicted annual total burned area from extreme fires and non-extreme fires increases from 2021 to 2050. Compared with the historical period, contributions of future climate change and anthropogenic activities to the annual total burned area are positive for non-extreme fires but negative for extreme fires due to reduced pasture expansion and deforestation. Therefore, mitigating climate change and implementing local sustainable land use strategies are crucial for restricting fires in the Brazilian Amazon. Earth and environmental sciences/Ecology/Fire ecology Earth and environmental sciences/Ecology/Ecosystem ecology Figures Figure 1 Figure 2 Figure 3 Figure 4 Main text The Amazon forests are pivotal in regulating the global terrestrial carbon cycle ( 1 , 2 ) and providing habitats for a rich diversity of species ( 3 ). These forests are facing a growing threat from intense fire disturbances, especially in areas where forests are fragmented and converted to other types of vegetation ( 4 , 5 ). Fire can alter vegetation structure and function ( 6 , 7 ) and affect the regional and global energy budget by changing surface albedo ( 8 ) and emitting aerosols and greenhouse gases ( 9 ), posing a significant threat to the valuable ecosystem services of tropical forests. In recent years, extreme fires have occurred more frequently and caused more severe damage to tropical forests than non-extreme fires; however, it remains unclear whether the drivers behind extreme and non-extreme fires differ ( 10 , 11 ). Understanding the drivers of fire disturbance while distinguishing between fire severity levels in the Brazilian Amazon region is therefore of great importance to understanding and protecting the world’s largest tropical forest from fire disturbances. Lightning-ignited wildfires and anthropogenic fires for farming expansion are two important sources of fire events in the Brazilian Amazon ( 12 , 13 ). According to satellite observations, there is a significant variation in both the size and intensity of these events ( 14 ). Previous studies have identified climatic and anthropogenic drivers of fires in the Brazilian Amazon, including fire foci ( 15 ), El Niño-Southern Oscillation events ( 16 – 20 ), logging ( 21 – 23 ), and expansion of roads ( 24 ), stockbreeding and agriculture ( 12 , 13 , 25 ). One experiment showed how a fire can evolve into an extreme one under suitable conditions ( 26 ), but the main drivers of extreme and non-extreme fires in the Brazilian Amazon remain unclear. Understanding these drivers is crucial for preventing non-extreme fires from escalating and causing severe damage to forests, and for providing guidance to policymakers to prevent and reduce fires in the face of global warming. Here, extreme fire is defined as 0.5°×0.5° grid cells with annual burned area exceeding the 90th percentile of the entire time series, while non-extreme fire includes 0.5°×0.5° grid cells with annual burned area greater than 0 but below the 90th percentile. In this study, we use machine learning models to build empirical relationships between climate, anthropogenic activities (represented by land use and land use change), and burned area in the Brazilian Amazon (see the workflow in Fig. S1 ). The spatial extent of the study region applied here is the Brazilian Amazon biome excluding the Cerrado (Methods). We first apply a random forest classification algorithm to categorize grid cells (0.5°×0.5°) into two types representing extreme fire and non-extreme fire, based on 4 different satellite-based burned area datasets covering the Brazilian Amazon during 1985 ~ 2020 (FireCCI51 ( 27 ), MCD64CMQ ( 28 ), MapBiomas Fire ( 29 ), and GABAM ( 30 )). We then use random forest quantile regression models to establish non-linear relationships between burned area, climate and anthropogenic variables in grid cells with extreme fires and non-extreme fires, respectively (Methods). The combined classification and regression models follow a previous study for reconstructing historical burned area ( 31 ). These random forest models are applied to predict future burned area in two different climate and land use change scenarios under Shared Socio-economic Pathways (SSPs) ( 32 ) and Representative Concentration Pathways (RCPs) ( 33 ). These are pairings of SSP1 & RCP2.6, SSP5 & RCP8.5. SSP1 & RCP2.6 represents a low-warming climate and more sustainable land use changes, whereas SSP5 & RCP8.5 represents a high-warming climate and less sustainable land use changes. Reproducing the observed burned area During 1985–2020, most observed burned area occurred on pasture, followed by forest and natural non-forest lands for both non-extreme and extreme fires (Fig. S2a, S2b). Deforestation-related (transition from forest to any other land use types) burned area accounts for about 10% of both non-extreme and extreme fires (Fig. S2). To reproduce the observed burned area, we perform feature selection (Methods) to retain the most influential input predictor variables, ultimately retaining 15 climate and anthropogenically-driven explanatory variables (Table S1 ) for the random forest models to predict extreme and non-extreme burned area in each 0.5°×0.5° grid cell. We leave out one year’s observed burned area data for validation and use the other years’ data to train the models and repeat the procedure for each year within the observation period. This ‘leave-one-year-out validation’ is thus focused on the ability of models to reproduce the inter-annual variations and spatial patterns of burned area. Random forest classification is first conducted on the grid cell level (0.5°×0.5°) with non-extreme fires or extreme fires, followed by random forest quantile regressions to predict burned area fraction of grid cells (0.5°×0.5°) for non-extreme fires and extreme fires, respectively (Fig. S1 ). The classification accuracy in the independent leave-one-year-out validation is around 80% (Fig. S3a), and area under curve (AUC) of non-extreme fires and extreme fires is around 0.9 (Fig. S3b) and 0.95 (Fig. S3c), respectively, indicating the good performance of the classification for the four different burned area datasets. For the evaluation of quantile regression models, we calculate the spatial coefficient of determination (R 2 ) and spatial root-mean-squared error (RMSE) between predictions and observations from independent validation data. The overall R 2 of non-extreme fires and extreme fires is around 0.50 (Fig. S3d), and the RMSE for non-extreme fires and extreme fires is around 0.02 (unitless, fraction in each 0.5°×0.5° grid cell, Fig. S3e) and 0.07 (Fig. S3f), respectively, indicating an adequate model performance for further analysis and prediction. In addition to the leave-one-year-out validation, quantile regression models are also validated against a random selection of 20% of data that are not used in model training (R 2 = 0.50 ~ 0.67 for non-extreme fires, R 2 = 0.58 ~ 0.72 for extreme fires, Fig. S4, S5). These show that the models for non-extreme fires exhibit slight overestimation of smaller burned area and underestimation of larger burned area, and the models for extreme fires manifest minor biases (Fig. S4). During 1985–2020, no significant linear trend in annual total burned area is detected for either non-extreme or extreme fires using the mean annual burned area of multiple observation datasets (Fig. S6a, S6b), but the increasing trend (0.021 million hectares yr − 2 , p-value < 0.01) for non-extreme fires is significant using the MapBiomas datasets (Fig. S6a). Spatially, 20.6% and 12.4% grid cells show significant positive and negative trends, respectively using the multiple-data mean observed burned area of all fires (Fig. S8). The model-predicted trends are generally consistent with the observed trends across the region (Fig. S9). The predicted area proportions of non-extreme and extreme fires are 40.4–47.8% and 52.2–59.6% (Fig. S7c-f), respectively, compared to 45.4–48.0% and 52.0-54.6% from observations (Fig. S6c-f). The observed and predicted spatial patterns of burned area in the Brazilian Amazon (Fig. 1 a, 1 c) exhibit overall consistency. Still, there are some spatial differences (Fig. 1 e, 1 f). For example, modelled burned area is overestimated in the southern regions (Fig. 1 e) versus observation-derived burned area data, however the latter also exhibit substantial inter-dataset variation in estimated burned area (Fig. 1 d). The modelled burned area fraction over the study region is significantly (p 0.8) and low spatial root mean square errors (RMSE ≤ 0.01) (Fig. 1 g-j). Systematic deviation is also minor, with the linear slope values ranging from 0.81 to 0.92 (Fig. 1 g-j). The modelled overall trend of annual total burned area in the Brazilian Amazon is consistent with observational estimates, although our predictions underestimate some peak years (Fig. 1 k), probably due to an insufficient sample size of large fires in the model training. Note that the standard deviation of annual total burned area across the four datasets is relatively large for both the observations and predictions (shaded area in Fig. 1 k), which is rooted from differences across datasets. Drivers of extreme and non-extreme fires The differentiation between extreme and non-extreme fires is driven mostly by anthropogenic factors related to land use and land use change (64.4 ± 5.2% contribution to explanatory distinction across the four burned area datasets, Fig. 2 a) rather than climatic factors (35.6 ± 5.2%) in the random forest classification (Fig. S1 ). In terms of individual factors distinguishing extreme and non-extreme fires, the most powerful are pasture fraction and forest fraction, followed by fire weather index and cumulative water deficit. Thus, the probability of extreme fires increases with annual increases in pasture fraction (Δpasture fraction, Fig. S1 1c) and the fire weather index, and decreases with annual increases in forest fraction (Δforest fraction, Fig. S1 1c). By contrast, the probability of non-extreme fires with land use change fractions and fire weather index is not as high as extreme fires (Fig. S1 1b). Both extreme and non-extreme burned area variation (random forest quantile regression, Fig. S1 ), were similarly driven by climatic variables (45.4% and 46.4%, respectively, Fig. 2 b, 2 c), while anthropogenic factors related to land use and land use change accounted for 54.6% and 53.6% of variation (Fig. 2 b, 2 c), respectively. This suggests that anthropogenic factors may play a slightly more important role in controlling burned area variation than climatic factors. Furthermore, among the anthropogenic factors, the impact of land use fraction (the current year’s land use status) is smaller for extreme fires (26.8%, Fig. 2 c) than non-extreme fires (34.2%, Fig. 2 b). By contrast, land use change fractions (i.e., land use fraction in the current year minus the previous year) exhibit more contribution in extreme fires (26.9%, Fig. 2 c) than non-extreme fires (20.4%, Fig. 2 b). Specifically, non-extreme fire burned area are driven mainly by (Fig. 2 a, 2 b), existing forest fraction, pasture fraction and fire weather index, which represent the current land use status and climate conditions. On the other hand, extreme fire burned area (Fig. 2 c) is driven mainly by the fire weather index, integrating climatic effects, such as temperature, humidity, wind speed and precipitation (Methods), on fire ignition and spread, followed by pasture and forest fraction change. This suggests that area changes in pasture and forest are more likely to contribute to the burned area of extreme fire compared to the current land use status of pasture and forest. However, correlations between burned area and land use change do not necessarily reflect direct causality between land use change and burned area variation, because fires can lead to forest loss, which is detected as land cover change from satellites. Addressing this, we conducted a causality test between land use change and burned area variation in extreme fire grid cells based on Convergence Cross Mapping (CCM) ( 34 ) (Methods). This analysis showed that in > 65% of grid cells with at least one fire occurrence (Fig. S12), land use change of forest and pasture drives burned area inter-annual variability, instead of the reverse. Daily maximum and minimum air temperature, representing daytime and nighttime temperature, respectively, are also important determinants of burned area variation. Daily maximum temperature controls evaporation and wetness and thus influence vegetation mortality, fuel load and flammability ( 35 , 36 ). Lower daily minimum temperature usually slows and extinguishes fires due to the greater heat consumption required for fire spread ( 37 ). Vapor pressure deficit (VPD), representing the atmosphere’s drying effects on fuels, controls fuel moisture ( 37 , 38 ), and is also an important factor in explaining fire spread for extreme fires (Fig. 2 c). In addition, drought events associated with a high VPD may also cause tree mortality, subsequently increasing fuel availability ( 39 ). We further analyzed the partial dependence of burned area on the two most important variables (Fig. 3 ) in each category of climate, land use and land use change, respectively (others in Fig. S11 and S13). While the rankings of variables may differ between non-extreme and extreme fires, their relationships with burned area exhibit some similarities. For both non-extreme (Fig. 3 c) and extreme fires (Fig. 3 g), the fire weather index shows a positive correlation, but the slope is relatively steeper for extreme fires. In addition, burned area for extreme fires increases more when the fire weather index exceeds a value of 5 (defined as “moderate fire danger” according to the Canadian Forest Fire Weather Index System ( 41 ); Fig. 3 g). However, burned area for non-extreme fires decrease with increasing daily minimum temperature (Fig. 3 e). One possible explanation is that non-forest areas tend to have lower nighttime temperatures than forested areas in the Brazilian Amazon ( 42 ), and fires in these non-forest regions usually burn larger areas. This could lead to the observed negative relationship between burned area and daily minimum temperature. In terms of anthropogenic variables, for both non-extreme (Fig. 3 b) and extreme fires (Fig. 3 k), increasing forest fraction tends to have a negative impact, while increasing pasture fraction has a positive impact on burned area (Fig. 3 a, 3 j). In pure-forest grid cells (forest fraction > 99%), very small burned area fractions are both observed and predicted by the model (Fig. S14). Moreover, land use conversion from forest to pasture (Fig. S15) leads to increased burned area across the fire severity spectrum (Fig. 3 d, 3 f, 3 h, 3 i). Projected burned area in the Brazilian Amazon We further predicted future burned area from 2021 to 2050 using the trained random forest models. As model input data forcing for these projections, we use future climate projection data from five climate models (GFDL-ESM4 ( 43 ), IPSL-CM6A-LR ( 44 ), MPI-ESM1-2-HR ( 45 ), MRI-ESM2-0 ( 46 ), and UKESM1-0-LL ( 47 )) in the Inter-Sectoral Impact Model Intercomparison Project 3b (hereafter ISIMIP3b) bias-adjusted climate data ( 48 ) and future land use data from a global dataset (Land-Use Harmonization version 2-future, hereafter LUHv2f ( 49 )) and a regional dataset for Brazil (from the Land Use and Land Cover Change Model to Brazil, hereafter LuccMEBR ( 50 )). These projections were performed for a low-warming scenario (SSP1 & RCP2.6, hereafter SSP126) and a high-warming scenario (SSP5 & RCP8.5, hereafter SSP585). The predicted future mean annual total burned area under SSP126 (4.6 ± 0.7 M ha yr − 1 , average of LUHv2f and LuccMEBR) is smaller than that under SSP585 (5.2 ± 0.7 M ha yr − 1 , average of LUHv2f and LuccMEBR; Fig. 4 ). Under SSP585, peak annual burned area of combined extreme and non-extreme fires occurs in the year 2043 at 6.7 ± 1.0 M ha yr − 1 (Fig. S16c), compared to the known historical maximum burned area of 8.1 M ha yr − 1 observed in 1997 (Fig. 1 k). According to model output, 15.7% of total time-integrated burned area is contributed by extreme fires under SSP126, compared to 23.2% under SSP585 (Fig. 4 ). This extreme fire contribution to burned area is much smaller than the corresponding historical contribution of 52.0%-54.6% (Fig. S6c-f) due to the decreased number of burned grid cells with projected extreme fires from the classification predictions (Fig. S17). Climate change increases non-extreme fire burned area but decreases extreme fire burned area in both SSP126 and SSP585 scenarios, with a larger impact in SSP585 (Fig. 4 ). By contrast, anthropogenic impacts increase extreme fire burned area under SSP585, but decrease it under SSP126, while increasing non-extreme burned area under both scenarios. Anthropogenic impacts are stronger using input data from LuccMEBR than LUHv2f, due to more intense land use change in the former (e.g., forest loss and pasture gain, Fig. S18, S20, S21). The contribution of climatic-anthropogenic interaction (Methods) is comparable to the anthropogenic effect, and it is positive for extreme fires but minor for non-extreme fires in both SSP126 and SSP585 scenarios (Fig. 4 ). This underscores the more important role of climatic-anthropogenic interactions in driving the burned area variation of extreme fires, compared to non-extreme fires. The predicted annual total burned area of all fires, extreme and non-extreme fires increase under both SSP126 and SSP585 scenarios (Fig. 4 c, 4 d). In terms of contributing factors, climate consistently exerts a positive influence on the increasing trends across all fire types under both scenarios (Table S4, Fig. S19). The burned area trends from anthropogenic effect are generally opposite between extreme and non-extreme fires under SSP126. The slopes of climatic-anthropogenic interaction effect for all fires, non-extreme and extreme fires are stronger in SSP585 than SSP126. Spatially, most large burned areas in future scenarios are located in the southeast and southwest Brazilian Amazon (Fig. S22) due to the larger projected land use changes in these regions (Fig. S20, S21). However, compared to the present-day pattern (Fig. 1 a, 1 c), future predicted burned area of all fires is smaller in these regions (Fig. S22) because of the decreased burned area contribution of extreme fires (Fig. S22, S23). On the other hand, we project increases in central Amazonian burned area mostly due to increased incidence of non-extreme fires (Fig. S22, S23). It should be noted that a discontinuity in annual burned area appears between 2020 and 2021 (Fig. S16), and the climatic effect on extreme fires is counterintuitively negative under both SSP126 and SSP585 (Fig. 4 a, 4 b). These issues likely stem from spatiotemporal inconsistencies (Methods) between historical observation-based and future model-predicted climate data (Supplementary Text 1.1, Fig. S40, S41), as well as discrepancies between present and future land use datasets (Supplementary Text 1.2, Fig. S38). We thus conducted sensitivity tests using different climate datasets and land use harmonization methods (Supplementary Texts 1 and 2). When isolating the impact of climate or land use, we found that inconsistencies in land use—rather than in climate—primarily explain the discontinuity in classifying extreme versus non-extreme grid cells from 2020 to 2021 (Supplementary Text 1). Therefore, the projected future burned area is more sensitive to inconsistencies in land use data than in climate data, and anthropogenic effects can influence the climatic effect on extreme fires through interaction terms (Fig. S45). Discussion Our study highlights the differences in the main drivers of burned area for extreme fires and non-extreme fires in the Brazilian Amazon using multiple long-term satellite-based burned area datasets. Area variation of non-extreme fires is primarily influenced by the land use status of pasture and forest (Fig. 2 b). Larger pasture fractions represent increased human activities in a region and a higher probability of ignitions ( 51 , 52 ). Compared to forests, fires also spread faster in pasture ( 51 , 53 , 54 ). This may partly explain why non-extreme burned area increases with pasture fraction (Fig. 3 c). However, increasing human land use may also cause landscape fragmentation ( 55 , 56 ), which may prevent a fire’s transition to an extreme fire. Indeed, extreme fire spread rates are driven by a combination of extreme weather conditions and recent land use change from forest to pasture (not the absolute land use fraction, Fig. 3 c). This may be because one side effect of clearing a forest for pasture is the conversion of large quantities of living biomass to dead wood and litter, thereby increasing fuel availability along with active ignitions ( 51 , 57 ). Deforestation also exposes interior forests to edge effects which can result in forest fragmentation ( 58 ), degradation ( 59 ) and increased fire risk ( 60 , 61 ). In addition, the causality test confirms that land use change drives burned area variation in most areas, rather than the opposite (Fig. S12). Thus, when weather conditions are favorable for fires, per a higher fire weather index value, degraded forests are vulnerable to extreme fires. We acknowledge the widespread uncertainties associated with analyzing the drivers of extreme and non-extreme fires and predicting future burned area in the Brazilian Amazon. Future climate and land use projection datasets inherently come with uncertainties. Furthermore, our models were trained using historical climate and land use observation datasets, and inconsistencies between these historical datasets and future projection datasets could introduce additional uncertainties. Despite these, the results of driver importance are consistent between different satellite-based burned area datasets, suggesting the robustness of our findings. We also made extensive tests of the different combinations of quantile predictions in reproducing the burned area for extreme fires (Methods, Fig. S24-S27), and validated the model using data from an independent year (Fig. S1 ). In addition, we tested different thresholds of defining extreme and non-extreme fires (i.e., 80th, 85th and 95th percentile), and the main drivers (Fig. S28, S29) are consistent with those using the 90th percentile (Fig. 2 ). Additionally, we separated forest, natural non-forest and pasture fires by overlapping burned area maps with land use maps (Fig. S2). Similar to Fig. 2 , pasture fraction and fire weather index are important for the burned area variations of non-extreme fires, and extreme fires are more driven by daily maximum temperature and fire weather index (Fig. S30). However, the importance of land use change to extreme fires decreases (Fig. S30) compared to not separating fires by land use types (Fig. 2 ), probably because it overlooks the burned area gradient across different land use types. Our study provides insight into the different drivers of non-extreme and extreme fires and predicts contributions of future climate change and land use change to burned area variations. Given the high uncertainty in mitigating climate change, which requires global efforts, implementing sustainable local land use strategies (e.g., controlling deforestation to pasture) becomes even more crucial for mitigating fires, especially extreme fires, in the Brazilian Amazon. Therefore, for policymakers and governing bodies in Brazil, it is crucial to implement measures that restrict anthropogenic fires for clearing natural vegetation and control excessive deforestation, particularly under extreme fire weather conditions. Methods Data preparation. We used four burned area products based on satellite observations (FireCCI51 ( 27 ); MCD64CMQ ( 28 ); MapBiomas Fire ( 29 ); GABAM ( 30 )). FireCCI51 ( 27 ) is a global burned area dataset available monthly from 2001 to 2020 with 0.25°×0.25° spatial resolution. MCD64CMQ ( 28 ) is one of the Moderate Resolution Imaging Spectroradiometer (MODIS) burned area products available monthly from 2001 to 2020 with 0.25°×0.25° spatial resolution. MapBiomas Fire ( 29 ) is a burned area product covering Brazil annually from 1985 to 2020 with 30 m spatial resolution. Global Annual Burned Area Maps (GABAM) ( 30 ) is a global burned area product annually from 1985 to 2020 with 30 m spatial resolution. Note that there are some missing years (1986, 1988, 1990, 1991, 1993, 1994, 1997 and 1999) in the GABAM dataset. The spatial extent of the Brazilian Amazon in this study is defined as the same region used in MapBiomas Brasil ( https://brasil.mapbiomas.org/infograficos/ ). The Multivariate ENSO Index is accessed from Physical Sciences Laboratory, National Oceanic and Atmospheric Administration ( https://psl.noaa.gov/enso/mei/#data ). We used 7 climatic variables and 8 anthropogenic variables related to land use and land use change together as input features to represent the climatic and anthropogenic effects on burned area. For the period of 1985–2020, climatic variables were derived from CRU JRA v2.2 ( 62 – 64 ), a global climate forcing dataset covering a time span from 1901 to 2020 with 6-hourly temporal resolution and 0.5°×0.5° spatial resolution. Land use related variables were extracted from the MapBiomas Collection 6 land use map ( 65 ), a land use product covering Brazil from 1985 to 2020 with 30 m spatial resolution, but it does not separate primary and secondary forest features. For the future period of 2021 ~ 2050, climatic variables were extracted from the Inter-Sectoral Impact Model Intercomparison Project 3b (ISIMIP3b) bias-adjusted climate data ( 48 ) of five different global climate models (GFDL-ESM4 ( 43 ), IPSL-CM6A-LR ( 44 ), MPI-ESM1-2-HR ( 45 ), MRI-ESM2-0 ( 46 ), and UKESM1-0-LL ( 47 )). This dataset, with a spatial resolution of 0.5°×0.5° and a time span of 2021 ~ 2100, has been bias corrected globally against W5E5 v2.0 ( 66 ) though, still manifesting spatiotemporal inconsistency with CRUJRA v2.2 in the Brazilian Amazon (Fig. S40, S41). We used the ISIMIP3b data for the SSP1&RCP2.6, SSP5&RCP8.5 scenarios to represent the low and high warming scenarios. Accordingly, we extracted future land use change data for these two scenarios from two land use projection datasets: Land-Use Harmonization version 2-future (hereafter LUHv2f ( 49 )) and LuccME/integrated surface process model (LuccME/INLAND) data (hereafter LuccMEBR ( 50 )). LUHv2f ( 49 ) is an annual global land use projection dataset from 2015 to 2100 with 0.25°×0.25° spatial resolution. LuccMEBR ( 50 ) is a land use projection dataset covering Brazil from 2015 to 2050 at a 5-year interval and about 10 km spatial resolution. LuccMEBR is generated by a regional LuccME modeling framework ( 67 ) with integration of global scenarios, thus it could be more representative on the regional scale than LUHv2f. LuccMEBR assumes reduced deforestation in the sustainable development scenario and no control of deforestation in the strong inequality scenario. All datasets were aggregated into 0.5°×0.5° grid cells on an annual scale. Seven climatic variables include daily maximum temperature, daily minimum temperature, precipitation, wind speed, vapor pressure deficit, cumulative water deficit ( 68 ), fire weather index ( 41 ). Since fires usually occur in the dry months, all climatic variables were calculated over “dry months” instead of all months within a year. The definition of dry months in the Brazilian Amazon (i.e., monthly precipitation less than 100 mm) follows Aragão et al., 2007 ( 68 ). Temperature, precipitation and wind speed were directly extracted from the climate datasets. Vapor pressure deficit was calculated using Tetens’ empirical formula ( 69 ) based on temperature, surface pressure, specific humidity. Fire weather index, as in Canadian Forest Fire Weather Index System ( 41 ), was calculated at the daily scale from temperature, relative humidity, wind speed and precipitation using ‘cffdrs’ package( 70 ) in R. The calculation of cumulative water deficit follows Aragão et al., 2007 ( 68 ) using the climate datasets mentioned above. Eight anthropogenic variables, represented by land use and land use change, include forest fraction, pasture fraction, natural non-forest vegetation fraction, cropland fraction, Δforest fraction, Δpasture fraction, Δnatural non-forest vegetation fraction, Δcropland fraction. Fraction refers to the area fraction of each land use type in a grid cell in the current year, and ‘Δfraction’ is changes in area fraction for a give land use type between two adjacent years. Collinearity tests of the 15 input features were conducted (Fig. S31). Model training and validation. In this study, we first conducted random forest classification to identify grid cells (0.5°×0.5°) with non-extreme and extreme fires. We define grid cells with extreme fires as grid cells with burned area exceeding the 90th percentile of the entire time series' observations. The other grid cells with burned area greater than 0 are thus treated as grid cells with non-extreme fires. Extreme fire is usually defined by a percentile threshold of fire radiative power, fire size or fire expansion speed ( 71 , 72 ). As sensitivity tests, we also tried other thresholds (80th, 85th and 95th percentile) to define the extreme fires (Fig. S28, S29). For grid cells with non-extreme or extreme fires, we performed random forest quantile regression of burned area separately. The output of quantile regressors is probability distribution, and quantiles of predictions can be retrieved from the distribution. Because the normal random forest algorithm is more conservative, it is not appropriate for predicting burned area of extreme fires due to the small sample number in the training dataset (Fig. S32). Therefore, we used a combination of different quantile predictions to best match the observations. Specifically, two or three output percentiles from 5th, 10th, …, 95th quantiles were randomly selected as a combination. Then the performance of the selected quantile combination was tested against the observations in the test dataset. Note that 20% of the samples were randomly split for testing, and the remaining 80% samples were used for training. The combination with the highest R 2 and the lowest RMSE was set as the final combination of output quantiles. The R 2 of all quantile combinations during this process are shown in Fig. S24 and S25 for non-extreme fires and in Fig. S26 and S27 for extreme fires. We also tried other machine learning methods such as normal random forest regression, gradient boosting regression, extreme gradient boosting regression and long-short-term memory neuron network regression, but their performances (Fig. S33) are not as good as the random forest quantile regression (Fig. 1 g-j). In addition to the 20% samples for testing mentioned above, the model was also validated independently year by year within the time span of burned area observations. For the data in a given year used for validation, data from the remaining years excluding this year were used to train the models, and predictions from the trained models were compared with the observed burned area in this year. Note that a significant positive linear relationship is detected between the R 2 of independent validations and the anomalies of annual burned area (Fig. S34, S35), suggesting our models perform better in years with larger positive burned area anomalies. Additionally, we also conducted feature selection with recursive feature elimination cross validation using the RFECV function in “scikit-learn” package ( 73 ) in Python, and the feature selection results are shown in Table S1 -S3, Fig. S36. The SHapley Additive exPlanations (SHAP) value is calculated with the “shap” package ( 40 ) using tree explainer in Python. Causality tests. Because land use change (LUC) may cause fires, and fires also result in land cover change, the direction of causality between LUC and burned area variation needs to be further confirmed. To prove whether LUC drives the inter-annual variability of burned area (IAV_BA) rather than the reverse, we applied Convergent Cross Mapping (CCM) ( 34 ) analysis to test the causality between LUC and IAV_BA. CCM is a causality identification method to distinguish causality from correlation based on non-linear state space reconstruction, which claim to be suitable for non-separable and weakly connected dynamic systems (e.g., ecosystems) ( 34 ). The CCM coefficient, ranging from − 1 to 1, quantifies the strength of causal influence, with larger absolute values indicating stronger causal effects. If the causal effect of A on B (denoted as A:B) is stronger than the reverse (B:A), A is recognized as the cause of B. The CCM analysis was conducted on the annual series of ever burned grid cells within the observation period based on the four burned area datasets (FireCCI, MODIS, GABAM and MapBiomas) (Fig. S37). In addition to burned area, area change of forest and pasture, as the two most important land use types according to the SHAP value ranking of machine learning models (Fig. 2 ), were taken into consideration. For each ever-burned grid cell, we extracted annual time series of the selected variables (burned area, ΔForest, and ΔPasture) over the whole observation period, and calculated the CCM coefficients using the “CCM” function in the “pyEDM” package ( 74 ) in Python. We set no embedding delay and fixed the prediction interval at 1 year, looping the embedding dimension from 1 to 5 to select the optimal parameter. If the CCM coefficient for ΔForest:burned area is larger than that for burned area:ΔForest, then ΔForest is considered to drive burned area change in that grid cell; the same approach was applied to ΔPasture and burned area. The CCM coefficient maps are shown in Fig. S12. When both ΔForest and ΔPasture influence burned area, we selected the larger absolute CCM coefficient (the third column of subplots in Fig. S12). Model prediction. After the model training and validation, we used the trained random forest models to predict future burned area during 2021–2050. Because both climate (CRUJRA and ISIMIP3b) and land use datasets (MapBiomas, LUHv2f, and LuccMEBR) were inconsistent between the historical period (1985–2020) and the future (2021–2050), we conducted several sensitivity tests using several harmonization methods to check the robustness of our results (Supplementary Text 1). For the climate datasets, we tested CRUJRA and ISIMIP3b respectively in model training (Supplementary Text 1.1, Fig. S40 and S41). For the land use datasets, land use types in all products were first reclassified to five categories: forest, non-forest natural vegetation, pasture, cropland and others. To harmonize these land use maps from the present day into the future, we used the interannual spatial changes recorded in each future dataset (2020–2050) and incrementally applied them to the 2020 MapBiomas baseline. Specifically, we calculated the spatial difference maps between two adjacent years from 2020 to 2050 and accumulated them year by year to the 2020 land use map of MapBiomas. In addition to this harmonization method (denoted as “spatial difference”), we also tested another two harmonization methods (see details in Supplementary Text 1.2). In the harmonization processes, if the harmonized fraction of one land use type exceeds 1 in a grid cell, it remains to be 1 and the fractions of other land use types are set to be 0; if it is less than 0, it is set as 0, and the deficit is allocated to other land use types proportionally to their fractions. We performed a set of predictions to separate the contributions of future climate and anthropogenic activities to burned area under the SSP126 and SSP585, including predictions using 1) fixed climate (the average state during 2011 ~ 2020) and land use (the same as 2020), 2) fixed climate but dynamic land use changes, 3) fixed land use but transient climate conditions and 4) transient climate conditions and dynamic land use changes. The anthropogenic effect is calculated from the difference between the first two scenarios, and climate effect is from the difference between the first and the third scenarios. The interaction between climate change and anthropogenic activities are calculated as the difference between the first and the fourth scenarios minus the climate effect and anthropogenic effect. Declarations Acknowledgements This study was supported by Yunnan Provincial Science and Technology Project at Southwest United Graduate School (grant number: 202302AO370001), the National Natural Science Foundation of China (grant number: 42175169), the Hainan Institute of National Park Research Program KY-23ZK01, and the CALIPSO (Carbon Loss In Plants, Soils and Oceans), funded through Schmidt Sciences LLC. SS, SB, PC were supported by ESA XFires (contract number 4000145351/24/I-LR). Data Availability FireCCI51 can be downloaded from https://geogra.uah.es/fire_cci/firecci51.php . MCD64CMQ can be downloaded from https://lpdaac.usgs.gov/products/mcd64a1v006/ . MapBiomas fire can be downloaded from https://brasil.mapbiomas.org/en/mapbiomas-fogo/ . Global Annual Burned Area Map (GABAM) can be downloaded from https://vapd.gitlab.io/post/gabam/ . MapBiomas land use maps can be downloaded from https://brasil.mapbiomas.org/en/colecoes-mapbiomas/ . Land-Use Harmonization version 2-future (LUH v2f) datasets can be downloaded from https://luh.umd.edu/data.shtml . LuccME/INLAND land-use scenarios for Brazil 2050 can be downloaded from https://zenodo.org/records/10611737 . CRUJRA v2.2 can be downloaded from https://catalogue.ceda.ac.uk/uuid/4bdf41fc10af4caaa489b14745c665a6 . ISIMIP3b IPSL output can be downloaded from https://www.isimip.org/gettingstarted/data-access/ . References Liu YY et al (2015) Recent reversal in loss of global terrestrial biomass. Nat Clim Change 5:470–474 Malhi Y et al (2008) Climate Change, Deforestation, and the Fate of the Amazon. Science 319:169–172 Barlow J et al (2016) Anthropogenic disturbance in tropical forests can double biodiversity loss from deforestation. Nature 535:144–147 Skole D, Tucker C (1993) Tropical Deforestation and Habitat Fragmentation in the Amazon: Satellite Data from 1978 to 1988. Science 260:1905–1910 Barlow J, Berenguer E, Carmenta R, França F (2020) Clarifying Amazonia's burning crisis. Glob Change Biol 26:319–321 Bond WJ, Woodward FI, Midgley GF (2005) The global distribution of ecosystems in a world without fire. New Phytol 165:525–538 Lasslop G, Brovkin V, Reick CH, Bathiany S, Kloster S (2016) Multiple stable states of tree cover in a global land surface model due to a fire-vegetation feedback. Geophys Res Lett 43:6324–6331 Randerson JT et al (2006) The Impact of Boreal Forest Fire on Climate Warming. Science 314:1130–1132 van der Werf GR et al (2010) Global fire emissions and the contribution of deforestation, savanna, forest, agricultural, and peat fires (1997–2009). Atmos Chem Phys 10:11707–11735 Aragão LEOC, Shimabukuro YE (2010) The Incidence of Fire in Amazonian Forests with Implications for REDD. Science 328:1275–1278 Mataveli G et al (2022) Record-breaking fires in the Brazilian Amazon associated with uncontrolled deforestation. Nat Ecol Evol 6:1792–1793 Pivello VR (2011) The Use of Fire in the Cerrado and Amazonian Rainforests of Brazil: Past and Present. Fire Ecol 7:24–39 Pendrill F et al (2022) Disentangling the numbers behind agriculture-driven tropical deforestation. Science 377:eabm9267 Andela N et al Tracking and classifying Amazon fire events in near real time. Sci Adv 8, eabd2713 da Silva Junior CA et al (2020) Persistent fire foci in all biomes undermine the Paris Agreement in Brazil. Sci Rep 10:16246 Chen Y et al (2011) Forecasting Fire Season Severity in South America Using Sea Surface Temperature Anomalies. Science 334:787–791 Fonseca MG et al (2017) Climatic and anthropogenic drivers of northern Amazon fires during the 2015–2016 El Niño event. Ecol Appl 27:2514–2527 Aragão LEOC et al (2018) 21st Century drought-related fires counteract the decline of Amazon deforestation carbon emissions. Nat Commun 9:536 Cai W et al (2020) Climate impacts of the El Niño–Southern Oscillation on South America. Nat Reviews Earth Environ 1:215–231 Wang Y, Huang P (2022) Potential fire risks in South America under anthropogenic forcing hidden by the Atlantic Multidecadal Oscillation. Nat Commun 13:2437 Nepstad DC et al (1999) Large-scale impoverishment of Amazonian forests by logging and fire. Nature 398:505–508 Fearnside PM (2005) Deforestation in Brazilian Amazonia: History, Rates, and Consequences. Conserv Biol 19:680–688 Balch JK, Nepstad DC, Curran LM (2009) In: Cochrane MA (ed) Tropical Fire Ecology: Climate Change, Land Use, and Ecosystem Dynamics. Springer Berlin Heidelberg, Berlin, Heidelberg, pp 481–502 Bowring SPK, Li W, Mouillot F, Rosan TM, Ciais P (2024) Road fragment edges enhance wildfire incidence and intensity, while suppressing global burned area. Nat Commun 15:9176 Sills J, Arruda D, Candido HG, Fonseca R (2019) Amazon fires threaten Brazil's agribusiness. Science 365:1387–1387 Brando PM et al (2014) Abrupt increases in Amazonian tree mortality due to drought–fire interactions. Proceedings of the National Academy of Sciences 111, 6347–6352 E. Chuviecoet al., Generation and analysis of a new global burned area product based on MODIS 250 m reflectance bands and thermal anomalies. Earth Syst. Sci. Data 10, 2015–2031 (2018) Giglio L, Boschetti L, Roy DP, Humber ML, Justice CO (2018) The Collection 6 MODIS burned area mapping algorithm and product. Remote Sens Environ 217:72–85 MapBiomas-Project (2021) Fire Scars (Collection 2 - MapBiomas Fire). (MapBiomas Project Long T et al (2019) 30 m Resolution Global Annual Burned Area Mapping Based on Landsat Images and Google Earth Engine. Remote Sens 11:489 Guo Z et al (2025) Reconstructed global monthly burned area maps from 1901 to 2020. Earth Syst. Sci. Data Discuss. 1–28 (2025) O’Neill BC et al (2014) A new scenario framework for climate change research: the concept of shared socioeconomic pathways. Clim Change 122:387–400 van Vuuren DP et al (2011) The representative concentration pathways: an overview. Clim Change 109:5 Sugihara G et al (2012) Detecting Causality in Complex Ecosystems. Science 338:496–500 Lima CHR, AghaKouchak A, Randerson JT (2018) Unraveling the Role of Temperature and Rainfall on Active Fires in the Brazilian Amazon Using a Nonlinear Poisson Model. J Geophys Research: Biogeosciences 123:117–128 Jain P, Castellanos-Acuna D, Coogan SCP, Abatzoglou JT, Flannigan MD (2022) Observed increases in extreme fire weather driven by atmospheric humidity and temperature. Nat Clim Change 12:63–70 Balch JK et al (2022) Warming weakens the night-time barrier to global fire. Nature 602:442–448 Nolan RH et al (2016) Predicting dead fine fuel moisture at regional scales using vapour pressure deficit from MODIS and gridded weather data. Remote Sens Environ 174:100–108 Grossiord C et al (2020) Plant responses to rising vapor pressure deficit. New Phytol 226:1550–1566 Lundberg SM et al (2020) From local explanations to global understanding with explainable AI for trees. Nat Mach Intell 2:56–67 Equations and FORTRAN program for the Canadian Forest Fire Weather Index System (1985) https://cfs.nrcan.gc.ca/publications?id=19973 ) Maeda EE et al (2021) Large-scale commodity agriculture exacerbates the climatic impacts of Amazonian deforestation. Proceedings of the National Academy of Sciences 118, e2023787118 Dunne JP et al (2020) The GFDL Earth System Model Version 4.1 (GFDL-ESM 4.1): Overall Coupled Model Description and Simulation Characteristics. Journal of Advances in Modeling Earth Systems 12, e2019MS002015 Boucher O et al (2020) Presentation and Evaluation of the IPSL-CM6A-LR Climate Model. J Adv Model Earth Syst 12:e2019MS002010 Gutjahr O et al (2019) Max Planck Institute Earth System Model (MPI-ESM1.2) for the High-Resolution Model Intercomparison Project (HighResMIP). Geosci Model Dev 12:3241–3281 Yukimoto S, The Meteorological Research Institute Earth System Model Version et al (2019) 2.0, MRI-ESM2.0: Description and Basic Evaluation of the Physical Component. J Meteorological Soc Japan Ser II 97:931–965 Sellar AA et al (2019) UKESM1: Description and Evaluation of the U.K. Earth System Model. J Adv Model Earth Syst 11:4513–4558 Lange S, Büchner M (2021) ISIMIP3b bias-adjusted atmospheric climate input data (v1.1). ISIMIP Repository Hurtt GC et al (2020) Harmonization of global land use change and management for the period 850–2100 (LUH2) for CMIP6. Geosci Model Dev 13:5425–5464 Silva Bezerra FG et al (2022) New land-use change scenarios for Brazil: Refining global SSPs with a regional spatially-explicit allocation model. PLoS ONE 17:e0256052 Cochrane MA (2003) Fire science for rainforests. Nature 421:913–919 Tyukavina A et al (2017) Types and rates of forest disturbance in Brazilian Legal Amazon, 2000–2013. Sci Adv 3:e1601047 Silvério DV et al (2013) Testing the Amazon savannization hypothesis: fire effects on invasion of a neotropical forest by native cerrado and exotic pasture grasses. Philosophical Trans Royal Soc B: Biol Sci 368:20120427 Cornwall W (2022) Fire Invasions. Around the world, flammable invasive grasses are increasing the risks of damaging wildfires. Science 377 Haddad NM et al (2015) Habitat fragmentation and its lasting impact on Earth’s ecosystems. Sci Adv 1:e1500052 Brinck K et al (2017) High resolution analysis of tropical forest fragmentation and its impact on the global carbon cycle. Nat Commun 8:14855 Brando PM, Oliveria-Santos C, Rocha W, Cury R, Coe MT (2016) Effects of experimental fuel additions on fire intensity and severity: unexpected carbon resilience of a neotropical forest. Glob Change Biol 22:2516–2525 Rosan TM et al (2022) Fragmentation-Driven Divergent Trends in Burned Area in Amazonia and Cerrado. Front Forests Global Change 5 Zhao Z et al (2021) Fire enhances forest degradation within forest edge zones in Africa. Nat Geosci 14:479–483 Cochrane MA (2001) Synergistic Interactions between Habitat Fragmentation and Fire in Evergreen Tropical Forests. Conserv Biol 15:1515–1521 Armenteras D, González TM, Retana J (2013) Forest fragmentation and edge influence on fire occurrence and intensity under different management types in Amazon forests. Biol Conserv 159:73–79 Harris I, Jones PD, Osborn TJ, Lister DH (2014) Updated high-resolution grids of monthly climatic observations – the CRU TS3.10 Dataset. Int J Climatol 34:623–642 Kobayashi S et al (2015) The JRA-55 Reanalysis: General Specifications and Basic Characteristics. J Meteorological Soc Japan Ser II 93:5–48 Harris I, Osborn TJ, Jones P, Lister D (2020) Version 4 of the CRU TS monthly high-resolution gridded multivariate climate dataset. Sci Data 7:109 MapBiomas-Project (2022) Collection 6 of the Annual Series of Land Use and Land Cover Maps of Brazil. (MapBiomas Project Cucchi M et al (2020) WFDE5: bias-adjusted ERA5 reanalysis data for impact studies. Earth Syst Sci Data 12:2097–2120 Aguiar AP, Soler AP, Assis L (2011) T., in GLP Land Use Transitions in South America: framing the present, preparing for a sustainable future Workshop Ilhabela Aragão LEOC et al (2007) Spatial patterns and fire response of recent Amazonian droughts. Geophys Res Lett 34 Tetens O (1930) Uber einige meteorologische begriffe. Z Geophys 6 Wang X et al (2017) cffdrs: an R package for the Canadian Forest Fire Danger Rating System. Ecol Processes 6:5 Bowman DMJS et al (2017) Human exposure and sensitivity to globally extreme wildfire events. Nat Ecol Evol 1:0058 Francisco Castro PM, Rego P, Fernandes C, Hoffman (2021) Fire Science . Springer Textbooks in Earth Sciences, Geography and EnvironmentSpringer Cham, ed. 1, pp. XXXVIII, 644 Pedregosa F et al (2011) Scikit-learn: Machine Learning in Python. J Mach Learn Res 12:2825–2830 Park J, GitHub (2019) https://github.com/SugiharaLab/pyEDM Additional Declarations There is NO Competing Interest. Supplementary Files SIAmazonBA250727.docx Supplementary text, tables and figures for Manuscript Cite Share Download PDF Status: Under Review Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7604250","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":521049452,"identity":"c22d8254-0e45-4a42-a236-3ca24969c843","order_by":0,"name":"Wei Li","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABA0lEQVRIiWNgGAWjYDACdsbGBwkGNnIMDIwNIAQCBvi1MDM2G3woSDPmIUELA5vkjA+HE3tAHKK0GBxmbpDmMTicvp//cNvDrzsOyzOwN2+TYKi5g0cLY4Mxj0F6bo9EYrux7JnDhg08x8okGI49w6nFDKglmcfAGqiFsU1asg3Ilcgxk2BsOIxXy2EeA+Z0Hv6DYC32DfJvCGppbJxh4JzAw5DYJvmx7XBigwQPfi32hxmbGT4YpBn23Ehsk2ZsS09u40krtkg4hluLZHv78x8Jf2zk2fuPP5P82WZt289+eOONDzW4taAAZh4gwQZiJRCnARiTP4hVOQpGwSgYBSMKAADzdFT9Qx7W4AAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0000-0003-2543-2558","institution":"Tsinghua University","correspondingAuthor":true,"prefix":"","firstName":"Wei","middleName":"","lastName":"Li","suffix":""},{"id":521049453,"identity":"39799547-5059-472c-9f5e-a3d77c00b671","order_by":1,"name":"Zhixuan Guo","email":"","orcid":"https://orcid.org/0009-0000-9996-6654","institution":"Tsinghua Univerity","correspondingAuthor":false,"prefix":"","firstName":"Zhixuan","middleName":"","lastName":"Guo","suffix":""},{"id":521049454,"identity":"75dcfbd8-5960-4f73-ac19-d6485d33b885","order_by":2,"name":"Philippe Ciais","email":"","orcid":"https://orcid.org/0000-0001-8560-4943","institution":"Laboratoire des Sciences du Climat et de l’Environnement, LSCE/IPSL, CEA-CNRS- UVSQ, Université Paris-Saclay, Gif-sur-Yvette, France","correspondingAuthor":false,"prefix":"","firstName":"Philippe","middleName":"","lastName":"Ciais","suffix":""},{"id":521049455,"identity":"dc40e096-2d6b-4302-895c-34c045367fc9","order_by":3,"name":"Stephen Sitch","email":"","orcid":"https://orcid.org/0000-0003-1821-8561","institution":"University of Exeter","correspondingAuthor":false,"prefix":"","firstName":"Stephen","middleName":"","lastName":"Sitch","suffix":""},{"id":521049456,"identity":"0fee724d-5f7c-4b6c-89d1-0fa6e2d8a0c1","order_by":4,"name":"Guido van der Werf","email":"","orcid":"https://orcid.org/0000-0001-9042-8630","institution":"Wageningen University \u0026 Research","correspondingAuthor":false,"prefix":"","firstName":"Guido","middleName":"van der","lastName":"Werf","suffix":""},{"id":521049457,"identity":"b18f4c92-5096-49b6-a517-2d54b6b23f45","order_by":5,"name":"Simon Bowring","email":"","orcid":"","institution":"Laboratoire des Sciences du Climat et de l'Environnement","correspondingAuthor":false,"prefix":"","firstName":"Simon","middleName":"","lastName":"Bowring","suffix":""},{"id":521049458,"identity":"57cb9a30-f444-410a-83cd-3944ebdc4c7a","order_by":6,"name":"Ana Bastos","email":"","orcid":"https://orcid.org/0000-0002-7368-7806","institution":"University of Leipzig","correspondingAuthor":false,"prefix":"","firstName":"Ana","middleName":"","lastName":"Bastos","suffix":""},{"id":521049459,"identity":"e6633d8b-83b3-40a0-9fb8-3484be6fbc7b","order_by":7,"name":"Dominic Fawcett","email":"","orcid":"","institution":"Swiss Federal Institute for Forest, Snow and Landscape Research WSL","correspondingAuthor":false,"prefix":"","firstName":"Dominic","middleName":"","lastName":"Fawcett","suffix":""},{"id":521049460,"identity":"66e8089d-7bdf-41d3-9fa2-fb7b5952dd89","order_by":8,"name":"Xiaomeng Du","email":"","orcid":"","institution":"Tsinghua University","correspondingAuthor":false,"prefix":"","firstName":"Xiaomeng","middleName":"","lastName":"Du","suffix":""},{"id":521049461,"identity":"b9e57601-28ac-4811-ae9a-cd5d69a425a4","order_by":9,"name":"Lei Zhu","email":"","orcid":"","institution":"Tsinghua University","correspondingAuthor":false,"prefix":"","firstName":"Lei","middleName":"","lastName":"Zhu","suffix":""},{"id":521049462,"identity":"b00d7520-f8b0-4211-8cb2-6744f38cb44a","order_by":10,"name":"Jiaying He","email":"","orcid":"https://orcid.org/0000-0002-6394-5218","institution":"Tsinghua University","correspondingAuthor":false,"prefix":"","firstName":"Jiaying","middleName":"","lastName":"He","suffix":""},{"id":521049463,"identity":"1558e900-fa7c-4b6b-8c2d-eee832732faf","order_by":11,"name":"Zhe Zhao","email":"","orcid":"","institution":"Taizhou University","correspondingAuthor":false,"prefix":"","firstName":"Zhe","middleName":"","lastName":"Zhao","suffix":""},{"id":521049464,"identity":"9b71e33b-062d-47f8-a239-1432c8e87dff","order_by":12,"name":"Nan Wang","email":"","orcid":"","institution":"Tsinghua University","correspondingAuthor":false,"prefix":"","firstName":"Nan","middleName":"","lastName":"Wang","suffix":""}],"badges":[],"createdAt":"2025-09-13 03:10:14","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7604250/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7604250/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":92384069,"identity":"3c2acaff-2edd-4117-b2ef-b3946856f113","added_by":"auto","created_at":"2025-09-29 06:53:23","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1274501,"visible":true,"origin":"","legend":"","description":"","filename":"MSAmazonBA250727.docx","url":"https://assets-eu.researchsquare.com/files/rs-7604250/v1/b04f258687d3bb62dbc9aafe.docx"},{"id":92383616,"identity":"c35d4a10-2c9c-4300-b020-8584b00fb0d4","added_by":"auto","created_at":"2025-09-29 06:45:23","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":12750,"visible":true,"origin":"","legend":"","description":"","filename":"NCOMMS2573122.json","url":"https://assets-eu.researchsquare.com/files/rs-7604250/v1/eb8563146bc0ae4fc250a527.json"},{"id":92383626,"identity":"8815c9ce-1955-47a1-8e03-e130b0266927","added_by":"auto","created_at":"2025-09-29 06:45:23","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":8854090,"visible":true,"origin":"","legend":"","description":"","filename":"SIAmazonBA250727.docx","url":"https://assets-eu.researchsquare.com/files/rs-7604250/v1/780ba844087c7cbc9a99184e.docx"},{"id":92383631,"identity":"d1b81325-e606-4e6a-ac15-544ef2e9be5c","added_by":"auto","created_at":"2025-09-29 06:45:23","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":140157,"visible":true,"origin":"","legend":"","description":"","filename":"NCOMMS25731220enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7604250/v1/2549bfba86d33b6474987d4c.xml"},{"id":92383622,"identity":"fe74f3d1-7b47-42b3-af6e-a456c993c6db","added_by":"auto","created_at":"2025-09-29 06:45:23","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":343818,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7604250/v1/5ce8a368d25ab073b4bbb706.png"},{"id":92384073,"identity":"73b94d34-0a7b-4ab2-a810-4a2f80627ff0","added_by":"auto","created_at":"2025-09-29 06:53:23","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":218258,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7604250/v1/3a9a78ecb3bf3a13064e51ce.png"},{"id":92383628,"identity":"4d462131-593a-4b43-a056-2270a2cbdeeb","added_by":"auto","created_at":"2025-09-29 06:45:23","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":238279,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7604250/v1/9b901654d55afda60452e504.png"},{"id":92383630,"identity":"04c0d760-89a8-451a-a5de-a6676e998f60","added_by":"auto","created_at":"2025-09-29 06:45:23","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":329081,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7604250/v1/cecfb493bef8ab2c9da8cb29.png"},{"id":92384076,"identity":"9ee0894a-a708-49fd-8d3b-d739e62b6fbd","added_by":"auto","created_at":"2025-09-29 06:53:23","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":76593,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7604250/v1/f98c4402bd916abcc62a844c.png"},{"id":92384072,"identity":"2d384f58-9e3d-45f6-be1b-af25f2549033","added_by":"auto","created_at":"2025-09-29 06:53:23","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":57457,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7604250/v1/0f8f475e995eb2be2b123d72.png"},{"id":92384075,"identity":"e563e25d-a4a7-41de-ade8-f5e498b1ecd3","added_by":"auto","created_at":"2025-09-29 06:53:23","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":49966,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7604250/v1/3f075ae942cf3df52370c359.png"},{"id":92383629,"identity":"a0bcf1b6-a361-4fee-8f32-6a12c477ffa6","added_by":"auto","created_at":"2025-09-29 06:45:23","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":96726,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7604250/v1/36505ab827f516f6d29f4405.png"},{"id":92383634,"identity":"7c553028-dbf2-4278-b693-1ca9d93cb2d2","added_by":"auto","created_at":"2025-09-29 06:45:23","extension":"xml","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":137131,"visible":true,"origin":"","legend":"","description":"","filename":"NCOMMS25731220structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7604250/v1/c339ef513d90f6d0c1630de6.xml"},{"id":92383633,"identity":"6fb16a16-dcd8-42af-a248-57330a60a6f7","added_by":"auto","created_at":"2025-09-29 06:45:23","extension":"html","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":151469,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7604250/v1/13b67632502fa9bb83d57469.html"},{"id":92384065,"identity":"c90a368a-98c4-47a6-bd6e-1ef5d2790e7d","added_by":"auto","created_at":"2025-09-29 06:53:23","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":343818,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eComparison between satellite observed and model predicted burned area in the Brazilian Amazon.\u003c/strong\u003e \u003cstrong\u003ea\u003c/strong\u003e,\u003cstrong\u003ec\u003c/strong\u003e,\u003cstrong\u003ee\u003c/strong\u003e, Mean annual burned area fraction in each 0.5°×0.5° grid cell during 2001-2020 averaged over the four burned area datasets for observation (\u003cstrong\u003ea\u003c/strong\u003e), prediction (\u003cstrong\u003ec\u003c/strong\u003e) and their difference (\u003cstrong\u003ee\u003c/strong\u003e). \u003cstrong\u003eb\u003c/strong\u003e,\u003cstrong\u003ed\u003c/strong\u003e,\u003cstrong\u003ef\u003c/strong\u003e, Standard deviation (SD) for the mean annual burned area fraction during 2001-2020 across the four burned area datasets. \u003cstrong\u003eg,h,i,j,\u003c/strong\u003e Scatter plots of mean annual burned area fractions between observation and prediction for the four burned area datasets: FireCCI (\u003cstrong\u003eg\u003c/strong\u003e), MODIS (\u003cstrong\u003eh\u003c/strong\u003e), GABAM (\u003cstrong\u003ei\u003c/strong\u003e), and MapBiomas (\u003cstrong\u003ej\u003c/strong\u003e). \u003cstrong\u003ek\u003c/strong\u003e, Time series of observed and predicted annual total burned area in the Brazilian Amazon, and the solid line and shaded area represent the average and standard deviation across the four burned area datasets. Note that years with SD=0 represent only MapBiomas is available. For the prediction of each year, the model is trained by data from all the other years. Years with annual mean Multivariate ENSO Index\u0026gt; 1.0 and \u0026lt; -1.0 are marked as EI Niño and La Niña, respectively (Methods).\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-7604250/v1/d18e939d6f37bb53bfb252f4.png"},{"id":92383618,"identity":"9dc7ed23-f78c-4948-954e-cfbabea5a1d9","added_by":"auto","created_at":"2025-09-29 06:45:23","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":245691,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDominant drivers of separating grid cells with non-extreme and extreme fires (random forest classification, a) and of predicting burned area for non-extreme (random forest regression, b) and extreme fires (c). \u003c/strong\u003eColors represent explanatory variables related to climate, land use, and land use change, respectively. Symbols indicate results for different burned area datasets (FireCCI, MODIS, GABAM and MapBiomas), and the cross and error bar indicate the mean and standard deviation across the four datasets. The insert ring denotes the contributions of the three groups of all the 15 input variables (Table S1-S3), but only the ten most important variables are shown individually here. The mean absolute SHapley Additive exPlanations (\u003cem\u003e40\u003c/em\u003e) (SHAP) value of all variables is shown in Fig. S10.\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-7604250/v1/6496bb286fc25f02e23e6fff.png"},{"id":92384066,"identity":"d7f866f7-e0a8-40af-a164-9dc39bf4e308","added_by":"auto","created_at":"2025-09-29 06:53:23","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":238279,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePartial dependence of representative variables (i.e., the two most important variables in each category of climate, land use and land use change, respectively) for non-extreme fire and extreme fire. \u003c/strong\u003eColors indicate different observation datasets of burned area (FireCCI, MODIS, GABAM and MapBiomas), and the dashed line and shaded area represent the average and standard deviation across the four burned area datasets. Note that the x-axis and y-axis scales are different for non-extreme fire and extreme fire. The partial dependence of variables in Fig. 2 is shown in Fig. S11 and S13.\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-7604250/v1/532b08449adb28c31d936db1.png"},{"id":92383619,"identity":"ca65d290-6dff-464b-a115-2dcb1202f3d0","added_by":"auto","created_at":"2025-09-29 06:45:23","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":307270,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePredicted mean annual burned area during 2021-2050 with contributions from anthropogenic activities and climate and time series of annual burned area for non-extreme and extreme fires under the SSP1 \u0026amp; RCP2.6 (a,c) and SSP5 \u0026amp; RCP8.5 (b,d) scenarios.\u003c/strong\u003e \u003cstrong\u003ea,b,\u003c/strong\u003e Predicted mean annual burned area during 2021-2050 with contributions from anthropogenic activities, climate and their interaction. The insert pie plots show the area fraction of extreme and non-extreme fires in the total burned area. Error bars indicate the interannual variability (i.e., the standard deviation of burned area in all years). \u003cstrong\u003ec,d,\u003c/strong\u003eTime series of annual burned area for non-extreme and extreme fires. Error bars indicate the standard deviation propagated across the four burned area datasets (FireCCI, MODIS, GABAM, and MapBiomas), the five climate models (GFDL-ESM4, IPSL-CM6A-LR, MPI-ESM1-2-HR, MRI-ESM2-0, and UKESM1-0-LL) in ISIMIP3b and the two land use datasets (LUHv2f and LuccMEBR). Slope and p-value of linear trends are annotated, and asterisk indicates a significant level with p-value \u0026lt; 0.05.\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-7604250/v1/b513533ac4b2f0c1373e69ec.png"},{"id":92384729,"identity":"7384dc4e-bb2c-468d-9ea7-2870c4ed63bf","added_by":"auto","created_at":"2025-09-29 07:01:24","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1843246,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7604250/v1/fc7849f7-20f2-4c1a-b67d-db49a70b6d5f.pdf"},{"id":92383624,"identity":"403d8515-abc5-484a-86aa-da6986c22625","added_by":"auto","created_at":"2025-09-29 06:45:23","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":8854090,"visible":true,"origin":"","legend":"Supplementary text, tables and figures for Manuscript","description":"","filename":"SIAmazonBA250727.docx","url":"https://assets-eu.researchsquare.com/files/rs-7604250/v1/d8f617171f74cb1d27a5845e.docx"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"Distinct drivers of extreme and non-extreme fires in the Brazilian Amazon","fulltext":[{"header":"Main text","content":"\u003cp\u003eThe Amazon forests are pivotal in regulating the global terrestrial carbon cycle (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) and providing habitats for a rich diversity of species (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e). These forests are facing a growing threat from intense fire disturbances, especially in areas where forests are fragmented and converted to other types of vegetation (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e). Fire can alter vegetation structure and function (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e) and affect the regional and global energy budget by changing surface albedo (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e) and emitting aerosols and greenhouse gases (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e), posing a significant threat to the valuable ecosystem services of tropical forests. In recent years, extreme fires have occurred more frequently and caused more severe damage to tropical forests than non-extreme fires; however, it remains unclear whether the drivers behind extreme and non-extreme fires differ (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e). Understanding the drivers of fire disturbance while distinguishing between fire severity levels in the Brazilian Amazon region is therefore of great importance to understanding and protecting the world\u0026rsquo;s largest tropical forest from fire disturbances.\u003c/p\u003e\u003cp\u003eLightning-ignited wildfires and anthropogenic fires for farming expansion are two important sources of fire events in the Brazilian Amazon (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e). According to satellite observations, there is a significant variation in both the size and intensity of these events (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e). Previous studies have identified climatic and anthropogenic drivers of fires in the Brazilian Amazon, including fire foci (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e), El Ni\u0026ntilde;o-Southern Oscillation events (\u003cspan additionalcitationids=\"CR17 CR18 CR19\" citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e), logging (\u003cspan additionalcitationids=\"CR22\" citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e), and expansion of roads (\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e), stockbreeding and agriculture (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e). One experiment showed how a fire can evolve into an extreme one under suitable conditions (\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e), but the main drivers of extreme and non-extreme fires in the Brazilian Amazon remain unclear. Understanding these drivers is crucial for preventing non-extreme fires from escalating and causing severe damage to forests, and for providing guidance to policymakers to prevent and reduce fires in the face of global warming.\u003c/p\u003e\u003cp\u003eHere, extreme fire is defined as 0.5\u0026deg;\u0026times;0.5\u0026deg; grid cells with annual burned area exceeding the 90th percentile of the entire time series, while non-extreme fire includes 0.5\u0026deg;\u0026times;0.5\u0026deg; grid cells with annual burned area greater than 0 but below the 90th percentile. In this study, we use machine learning models to build empirical relationships between climate, anthropogenic activities (represented by land use and land use change), and burned area in the Brazilian Amazon (see the workflow in Fig. \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e). The spatial extent of the study region applied here is the Brazilian Amazon biome excluding the Cerrado (Methods). We first apply a random forest classification algorithm to categorize grid cells (0.5\u0026deg;\u0026times;0.5\u0026deg;) into two types representing extreme fire and non-extreme fire, based on 4 different satellite-based burned area datasets covering the Brazilian Amazon during 1985\u0026thinsp;~\u0026thinsp;2020 (FireCCI51 (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e), MCD64CMQ (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e), MapBiomas Fire (\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e), and GABAM (\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e)).\u003c/p\u003e\u003cp\u003eWe then use random forest quantile regression models to establish non-linear relationships between burned area, climate and anthropogenic variables in grid cells with extreme fires and non-extreme fires, respectively (Methods). The combined classification and regression models follow a previous study for reconstructing historical burned area (\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e). These random forest models are applied to predict future burned area in two different climate and land use change scenarios under Shared Socio-economic Pathways (SSPs) (\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e) and Representative Concentration Pathways (RCPs) (\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e). These are pairings of SSP1 \u0026amp; RCP2.6, SSP5 \u0026amp; RCP8.5. SSP1 \u0026amp; RCP2.6 represents a low-warming climate and more sustainable land use changes, whereas SSP5 \u0026amp; RCP8.5 represents a high-warming climate and less sustainable land use changes.\u003c/p\u003e"},{"header":"Reproducing the observed burned area","content":"\u003cp\u003eDuring 1985–2020, most observed burned area occurred on pasture, followed by forest and natural non-forest lands for both non-extreme and extreme fires (Fig. S2a, S2b). Deforestation-related (transition from forest to any other land use types) burned area accounts for about 10% of both non-extreme and extreme fires (Fig. S2). To reproduce the observed burned area, we perform feature selection (Methods) to retain the most influential input predictor variables, ultimately retaining 15 climate and anthropogenically-driven explanatory variables (Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e) for the random forest models to predict extreme and non-extreme burned area in each 0.5°×0.5° grid cell. We leave out one year’s observed burned area data for validation and use the other years’ data to train the models and repeat the procedure for each year within the observation period. This ‘leave-one-year-out validation’ is thus focused on the ability of models to reproduce the inter-annual variations and spatial patterns of burned area. Random forest classification is first conducted on the grid cell level (0.5°×0.5°) with non-extreme fires or extreme fires, followed by random forest quantile regressions to predict burned area fraction of grid cells (0.5°×0.5°) for non-extreme fires and extreme fires, respectively (Fig. \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eThe classification accuracy in the independent leave-one-year-out validation is around 80% (Fig. S3a), and area under curve (AUC) of non-extreme fires and extreme fires is around 0.9 (Fig. S3b) and 0.95 (Fig. S3c), respectively, indicating the good performance of the classification for the four different burned area datasets. For the evaluation of quantile regression models, we calculate the spatial coefficient of determination (R\u003csup\u003e2\u003c/sup\u003e) and spatial root-mean-squared error (RMSE) between predictions and observations from independent validation data. The overall R\u003csup\u003e2\u003c/sup\u003e of non-extreme fires and extreme fires is around 0.50 (Fig. S3d), and the RMSE for non-extreme fires and extreme fires is around 0.02 (unitless, fraction in each 0.5°×0.5° grid cell, Fig. S3e) and 0.07 (Fig. S3f), respectively, indicating an adequate model performance for further analysis and prediction. In addition to the leave-one-year-out validation, quantile regression models are also validated against a random selection of 20% of data that are not used in model training (R\u003csup\u003e2\u003c/sup\u003e = 0.50 ~ 0.67 for non-extreme fires, R\u003csup\u003e2\u003c/sup\u003e = 0.58 ~ 0.72 for extreme fires, Fig. S4, S5). These show that the models for non-extreme fires exhibit slight overestimation of smaller burned area and underestimation of larger burned area, and the models for extreme fires manifest minor biases (Fig. S4).\u003c/p\u003e\u003cp\u003eDuring 1985–2020, no significant linear trend in annual total burned area is detected for either non-extreme or extreme fires using the mean annual burned area of multiple observation datasets (Fig. S6a, S6b), but the increasing trend (0.021\u0026nbsp;million hectares yr\u003csup\u003e− 2\u003c/sup\u003e, p-value \u0026lt; 0.01) for non-extreme fires is significant using the MapBiomas datasets (Fig. S6a). Spatially, 20.6% and 12.4% grid cells show significant positive and negative trends, respectively using the multiple-data mean observed burned area of all fires (Fig. S8). The model-predicted trends are generally consistent with the observed trends across the region (Fig. S9). The predicted area proportions of non-extreme and extreme fires are 40.4–47.8% and 52.2–59.6% (Fig. S7c-f), respectively, compared to 45.4–48.0% and 52.0-54.6% from observations (Fig. S6c-f). The observed and predicted spatial patterns of burned area in the Brazilian Amazon (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ea, \u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ec) exhibit overall consistency. Still, there are some spatial differences (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ee, \u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ef). For example, modelled burned area is overestimated in the southern regions (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ee) versus observation-derived burned area data, however the latter also exhibit substantial inter-dataset variation in estimated burned area (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ed). The modelled burned area fraction over the study region is significantly (p \u0026lt; 0.01) correlated with the observational estimates across all grid cells, with high spatial R\u003csup\u003e2\u003c/sup\u003e values (\u0026gt; 0.8) and low spatial root mean square errors (RMSE ≤ 0.01) (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eg-j). Systematic deviation is also minor, with the linear slope values ranging from 0.81 to 0.92 (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eg-j). The modelled overall trend of annual total burned area in the Brazilian Amazon is consistent with observational estimates, although our predictions underestimate some peak years (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ek), probably due to an insufficient sample size of large fires in the model training. Note that the standard deviation of annual total burned area across the four datasets is relatively large for both the observations and predictions (shaded area in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ek), which is rooted from differences across datasets.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e"},{"header":"Drivers of extreme and non-extreme fires","content":"\u003cp\u003eThe differentiation between extreme and non-extreme fires is driven mostly by anthropogenic factors related to land use and land use change (64.4 ± 5.2% contribution to explanatory distinction across the four burned area datasets, Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea) rather than climatic factors (35.6 ± 5.2%) in the random forest classification (Fig. \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e). In terms of individual factors distinguishing extreme and non-extreme fires, the most powerful are pasture fraction and forest fraction, followed by fire weather index and cumulative water deficit. Thus, the probability of extreme fires increases with annual increases in pasture fraction (Δpasture fraction, Fig. \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e1c) and the fire weather index, and decreases with annual increases in forest fraction (Δforest fraction, Fig. \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e1c). By contrast, the probability of non-extreme fires with land use change fractions and fire weather index is not as high as extreme fires (Fig. \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e1b).\u003c/p\u003e\u003cp\u003eBoth extreme and non-extreme burned area variation (random forest quantile regression, Fig. \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e), were similarly driven by climatic variables (45.4% and 46.4%, respectively, Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb, \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec), while anthropogenic factors related to land use and land use change accounted for 54.6% and 53.6% of variation (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb, \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec), respectively. This suggests that anthropogenic factors may play a slightly more important role in controlling burned area variation than climatic factors. Furthermore, among the anthropogenic factors, the impact of land use fraction (the current year’s land use status) is smaller for extreme fires (26.8%, Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec) than non-extreme fires (34.2%, Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb). By contrast, land use change fractions (i.e., land use fraction in the current year minus the previous year) exhibit more contribution in extreme fires (26.9%, Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec) than non-extreme fires (20.4%, Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb).\u003c/p\u003e\u003cp\u003eSpecifically, non-extreme fire burned area are driven mainly by (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea, \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb), existing forest fraction, pasture fraction and fire weather index, which represent the current land use status and climate conditions. On the other hand, extreme fire burned area (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec) is driven mainly by the fire weather index, integrating climatic effects, such as temperature, humidity, wind speed and precipitation (Methods), on fire ignition and spread, followed by pasture and forest fraction change. This suggests that area changes in pasture and forest are more likely to contribute to the burned area of extreme fire compared to the current land use status of pasture and forest.\u003c/p\u003e\u003cp\u003eHowever, correlations between burned area and land use change do not necessarily reflect direct causality between land use change and burned area variation, because fires can lead to forest loss, which is detected as land cover change from satellites. Addressing this, we conducted a causality test between land use change and burned area variation in extreme fire grid cells based on Convergence Cross Mapping (CCM) (\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e) (Methods). This analysis showed that in \u0026gt; 65% of grid cells with at least one fire occurrence (Fig. S12), land use change of forest and pasture drives burned area inter-annual variability, instead of the reverse. Daily maximum and minimum air temperature, representing daytime and nighttime temperature, respectively, are also important determinants of burned area variation. Daily maximum temperature controls evaporation and wetness and thus influence vegetation mortality, fuel load and flammability (\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e, \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e). Lower daily minimum temperature usually slows and extinguishes fires due to the greater heat consumption required for fire spread (\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e). Vapor pressure deficit (VPD), representing the atmosphere’s drying effects on fuels, controls fuel moisture (\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e), and is also an important factor in explaining fire spread for extreme fires (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec). In addition, drought events associated with a high VPD may also cause tree mortality, subsequently increasing fuel availability (\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eWe further analyzed the partial dependence of burned area on the two most important variables (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e) in each category of climate, land use and land use change, respectively (others in Fig. S11 and S13). While the rankings of variables may differ between non-extreme and extreme fires, their relationships with burned area exhibit some similarities. For both non-extreme (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ec) and extreme fires (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eg), the fire weather index shows a positive correlation, but the slope is relatively steeper for extreme fires. In addition, burned area for extreme fires increases more when the fire weather index exceeds a value of 5 (defined as “moderate fire danger” according to the Canadian Forest Fire Weather Index System (\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e); Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eg). However, burned area for non-extreme fires decrease with increasing daily minimum temperature (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ee). One possible explanation is that non-forest areas tend to have lower nighttime temperatures than forested areas in the Brazilian Amazon (\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e), and fires in these non-forest regions usually burn larger areas. This could lead to the observed negative relationship between burned area and daily minimum temperature. In terms of anthropogenic variables, for both non-extreme (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb) and extreme fires (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ek), increasing forest fraction tends to have a negative impact, while increasing pasture fraction has a positive impact on burned area (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea, \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ej). In pure-forest grid cells (forest fraction \u0026gt; 99%), very small burned area fractions are both observed and predicted by the model (Fig. S14). Moreover, land use conversion from forest to pasture (Fig. S15) leads to increased burned area across the fire severity spectrum (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ed, \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ef, \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eh, \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ei).\u003c/p\u003e"},{"header":"Projected burned area in the Brazilian Amazon","content":"\u003cp\u003eWe further predicted future burned area from 2021 to 2050 using the trained random forest models. As model input data forcing for these projections, we use future climate projection data from five climate models (GFDL-ESM4 (\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e), IPSL-CM6A-LR (\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e), MPI-ESM1-2-HR (\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e), MRI-ESM2-0 (\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e), and UKESM1-0-LL (\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e)) in the Inter-Sectoral Impact Model Intercomparison Project 3b (hereafter ISIMIP3b) bias-adjusted climate data (\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e) and future land use data from a global dataset (Land-Use Harmonization version 2-future, hereafter LUHv2f (\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e)) and a regional dataset for Brazil (from the Land Use and Land Cover Change Model to Brazil, hereafter LuccMEBR (\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e)). These projections were performed for a low-warming scenario (SSP1 \u0026amp; RCP2.6, hereafter SSP126) and a high-warming scenario (SSP5 \u0026amp; RCP8.5, hereafter SSP585).\u003c/p\u003e\u003cp\u003eThe predicted future mean annual total burned area under SSP126 (4.6\u0026thinsp;\u0026plusmn;\u0026thinsp;0.7 M ha yr\u003csup\u003e\u0026minus;\u0026thinsp;1\u003c/sup\u003e, average of LUHv2f and LuccMEBR) is smaller than that under SSP585 (5.2\u0026thinsp;\u0026plusmn;\u0026thinsp;0.7 M ha yr\u003csup\u003e\u0026minus;\u0026thinsp;1\u003c/sup\u003e, average of LUHv2f and LuccMEBR; Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). Under SSP585, peak annual burned area of combined extreme and non-extreme fires occurs in the year 2043 at 6.7\u0026thinsp;\u0026plusmn;\u0026thinsp;1.0 M ha yr\u003csup\u003e\u0026minus;\u0026thinsp;1\u003c/sup\u003e (Fig. S16c), compared to the known historical maximum burned area of 8.1 M ha yr\u003csup\u003e\u0026minus;\u0026thinsp;1\u003c/sup\u003e observed in 1997 (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ek). According to model output, 15.7% of total time-integrated burned area is contributed by extreme fires under SSP126, compared to 23.2% under SSP585 (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). This extreme fire contribution to burned area is much smaller than the corresponding historical contribution of 52.0%-54.6% (Fig. S6c-f) due to the decreased number of burned grid cells with projected extreme fires from the classification predictions (Fig. S17).\u003c/p\u003e\u003cp\u003eClimate change increases non-extreme fire burned area but decreases extreme fire burned area in both SSP126 and SSP585 scenarios, with a larger impact in SSP585 (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). By contrast, anthropogenic impacts increase extreme fire burned area under SSP585, but decrease it under SSP126, while increasing non-extreme burned area under both scenarios. Anthropogenic impacts are stronger using input data from LuccMEBR than LUHv2f, due to more intense land use change in the former (e.g., forest loss and pasture gain, Fig. S18, S20, S21). The contribution of climatic-anthropogenic interaction (Methods) is comparable to the anthropogenic effect, and it is positive for extreme fires but minor for non-extreme fires in both SSP126 and SSP585 scenarios (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). This underscores the more important role of climatic-anthropogenic interactions in driving the burned area variation of extreme fires, compared to non-extreme fires.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe predicted annual total burned area of all fires, extreme and non-extreme fires increase under both SSP126 and SSP585 scenarios (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ec, \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ed). In terms of contributing factors, climate consistently exerts a positive influence on the increasing trends across all fire types under both scenarios (Table S4, Fig. S19). The burned area trends from anthropogenic effect are generally opposite between extreme and non-extreme fires under SSP126. The slopes of climatic-anthropogenic interaction effect for all fires, non-extreme and extreme fires are stronger in SSP585 than SSP126.\u003c/p\u003e\u003cp\u003eSpatially, most large burned areas in future scenarios are located in the southeast and southwest Brazilian Amazon (Fig. S22) due to the larger projected land use changes in these regions (Fig. S20, S21). However, compared to the present-day pattern (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ea, \u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ec), future predicted burned area of all fires is smaller in these regions (Fig. S22) because of the decreased burned area contribution of extreme fires (Fig. S22, S23). On the other hand, we project increases in central Amazonian burned area mostly due to increased incidence of non-extreme fires (Fig. S22, S23).\u003c/p\u003e\u003cp\u003eIt should be noted that a discontinuity in annual burned area appears between 2020 and 2021 (Fig. S16), and the climatic effect on extreme fires is counterintuitively negative under both SSP126 and SSP585 (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea, \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb). These issues likely stem from spatiotemporal inconsistencies (Methods) between historical observation-based and future model-predicted climate data (Supplementary Text 1.1, Fig. S40, S41), as well as discrepancies between present and future land use datasets (Supplementary Text 1.2, Fig. S38). We thus conducted sensitivity tests using different climate datasets and land use harmonization methods (Supplementary Texts 1 and 2). When isolating the impact of climate or land use, we found that inconsistencies in land use\u0026mdash;rather than in climate\u0026mdash;primarily explain the discontinuity in classifying extreme versus non-extreme grid cells from 2020 to 2021 (Supplementary Text 1). Therefore, the projected future burned area is more sensitive to inconsistencies in land use data than in climate data, and anthropogenic effects can influence the climatic effect on extreme fires through interaction terms (Fig. S45).\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eOur study highlights the differences in the main drivers of burned area for extreme fires and non-extreme fires in the Brazilian Amazon using multiple long-term satellite-based burned area datasets. Area variation of non-extreme fires is primarily influenced by the land use status of pasture and forest (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb). Larger pasture fractions represent increased human activities in a region and a higher probability of ignitions (\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e, \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e). Compared to forests, fires also spread faster in pasture (\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e, \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e, \u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e). This may partly explain why non-extreme burned area increases with pasture fraction (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ec). However, increasing human land use may also cause landscape fragmentation (\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e, \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e), which may prevent a fire\u0026rsquo;s transition to an extreme fire. Indeed, extreme fire spread rates are driven by a combination of extreme weather conditions and recent land use change from forest to pasture (not the absolute land use fraction, Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ec). This may be because one side effect of clearing a forest for pasture is the conversion of large quantities of living biomass to dead wood and litter, thereby increasing fuel availability along with active ignitions (\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e, \u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e). Deforestation also exposes interior forests to edge effects which can result in forest fragmentation (\u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e), degradation (\u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e) and increased fire risk (\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e, \u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e). In addition, the causality test confirms that land use change drives burned area variation in most areas, rather than the opposite (Fig. S12). Thus, when weather conditions are favorable for fires, per a higher fire weather index value, degraded forests are vulnerable to extreme fires.\u003c/p\u003e\u003cp\u003eWe acknowledge the widespread uncertainties associated with analyzing the drivers of extreme and non-extreme fires and predicting future burned area in the Brazilian Amazon. Future climate and land use projection datasets inherently come with uncertainties. Furthermore, our models were trained using historical climate and land use observation datasets, and inconsistencies between these historical datasets and future projection datasets could introduce additional uncertainties. Despite these, the results of driver importance are consistent between different satellite-based burned area datasets, suggesting the robustness of our findings. We also made extensive tests of the different combinations of quantile predictions in reproducing the burned area for extreme fires (Methods, Fig. S24-S27), and validated the model using data from an independent year (Fig. \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e). In addition, we tested different thresholds of defining extreme and non-extreme fires (i.e., 80th, 85th and 95th percentile), and the main drivers (Fig. S28, S29) are consistent with those using the 90th percentile (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). Additionally, we separated forest, natural non-forest and pasture fires by overlapping burned area maps with land use maps (Fig. S2). Similar to Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, pasture fraction and fire weather index are important for the burned area variations of non-extreme fires, and extreme fires are more driven by daily maximum temperature and fire weather index (Fig. S30). However, the importance of land use change to extreme fires decreases (Fig. S30) compared to not separating fires by land use types (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e), probably because it overlooks the burned area gradient across different land use types.\u003c/p\u003e\u003cp\u003eOur study provides insight into the different drivers of non-extreme and extreme fires and predicts contributions of future climate change and land use change to burned area variations. Given the high uncertainty in mitigating climate change, which requires global efforts, implementing sustainable local land use strategies (e.g., controlling deforestation to pasture) becomes even more crucial for mitigating fires, especially extreme fires, in the Brazilian Amazon. Therefore, for policymakers and governing bodies in Brazil, it is crucial to implement measures that restrict anthropogenic fires for clearing natural vegetation and control excessive deforestation, particularly under extreme fire weather conditions.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cb\u003eData preparation.\u003c/b\u003e We used four burned area products based on satellite observations (FireCCI51 (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e); MCD64CMQ (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e); MapBiomas Fire (\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e); GABAM (\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e)). FireCCI51 (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e) is a global burned area dataset available monthly from 2001 to 2020 with 0.25\u0026deg;\u0026times;0.25\u0026deg; spatial resolution. MCD64CMQ (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e) is one of the Moderate Resolution Imaging Spectroradiometer (MODIS) burned area products available monthly from 2001 to 2020 with 0.25\u0026deg;\u0026times;0.25\u0026deg; spatial resolution. MapBiomas Fire (\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e) is a burned area product covering Brazil annually from 1985 to 2020 with 30 m spatial resolution. Global Annual Burned Area Maps (GABAM) (\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e) is a global burned area product annually from 1985 to 2020 with 30 m spatial resolution. Note that there are some missing years (1986, 1988, 1990, 1991, 1993, 1994, 1997 and 1999) in the GABAM dataset. The spatial extent of the Brazilian Amazon in this study is defined as the same region used in MapBiomas Brasil (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://brasil.mapbiomas.org/infograficos/\u003c/span\u003e\u003cspan address=\"https://brasil.mapbiomas.org/infograficos/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). The Multivariate ENSO Index is accessed from Physical Sciences Laboratory, National Oceanic and Atmospheric Administration (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://psl.noaa.gov/enso/mei/#data\u003c/span\u003e\u003cspan address=\"https://psl.noaa.gov/enso/mei/#data\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eWe used 7 climatic variables and 8 anthropogenic variables related to land use and land use change together as input features to represent the climatic and anthropogenic effects on burned area. For the period of 1985\u0026ndash;2020, climatic variables were derived from CRU JRA v2.2 (\u003cspan additionalcitationids=\"CR63\" citationid=\"CR62\" class=\"CitationRef\"\u003e62\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e64\u003c/span\u003e), a global climate forcing dataset covering a time span from 1901 to 2020 with 6-hourly temporal resolution and 0.5\u0026deg;\u0026times;0.5\u0026deg; spatial resolution. Land use related variables were extracted from the MapBiomas Collection 6 land use map (\u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e65\u003c/span\u003e), a land use product covering Brazil from 1985 to 2020 with 30 m spatial resolution, but it does not separate primary and secondary forest features. For the future period of 2021\u0026thinsp;~\u0026thinsp;2050, climatic variables were extracted from the Inter-Sectoral Impact Model Intercomparison Project 3b (ISIMIP3b) bias-adjusted climate data (\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e) of five different global climate models (GFDL-ESM4 (\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e), IPSL-CM6A-LR (\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e), MPI-ESM1-2-HR (\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e), MRI-ESM2-0 (\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e), and UKESM1-0-LL (\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e)). This dataset, with a spatial resolution of 0.5\u0026deg;\u0026times;0.5\u0026deg; and a time span of 2021\u0026thinsp;~\u0026thinsp;2100, has been bias corrected globally against W5E5 v2.0 (\u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e66\u003c/span\u003e) though, still manifesting spatiotemporal inconsistency with CRUJRA v2.2 in the Brazilian Amazon (Fig. S40, S41). We used the ISIMIP3b data for the SSP1\u0026amp;RCP2.6, SSP5\u0026amp;RCP8.5 scenarios to represent the low and high warming scenarios. Accordingly, we extracted future land use change data for these two scenarios from two land use projection datasets: Land-Use Harmonization version 2-future (hereafter LUHv2f (\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e)) and LuccME/integrated surface process model (LuccME/INLAND) data (hereafter LuccMEBR (\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e)). LUHv2f (\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e) is an annual global land use projection dataset from 2015 to 2100 with 0.25\u0026deg;\u0026times;0.25\u0026deg; spatial resolution. LuccMEBR (\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e) is a land use projection dataset covering Brazil from 2015 to 2050 at a 5-year interval and about 10 km spatial resolution. LuccMEBR is generated by a regional LuccME modeling framework (\u003cspan citationid=\"CR67\" class=\"CitationRef\"\u003e67\u003c/span\u003e) with integration of global scenarios, thus it could be more representative on the regional scale than LUHv2f. LuccMEBR assumes reduced deforestation in the sustainable development scenario and no control of deforestation in the strong inequality scenario. All datasets were aggregated into 0.5\u0026deg;\u0026times;0.5\u0026deg; grid cells on an annual scale.\u003c/p\u003e\u003cp\u003eSeven climatic variables include daily maximum temperature, daily minimum temperature, precipitation, wind speed, vapor pressure deficit, cumulative water deficit (\u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e68\u003c/span\u003e), fire weather index (\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e). Since fires usually occur in the dry months, all climatic variables were calculated over \u0026ldquo;dry months\u0026rdquo; instead of all months within a year. The definition of dry months in the Brazilian Amazon (i.e., monthly precipitation less than 100 mm) follows Arag\u0026atilde;o et al., 2007 (\u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e68\u003c/span\u003e). Temperature, precipitation and wind speed were directly extracted from the climate datasets. Vapor pressure deficit was calculated using Tetens\u0026rsquo; empirical formula (\u003cspan citationid=\"CR69\" class=\"CitationRef\"\u003e69\u003c/span\u003e) based on temperature, surface pressure, specific humidity. Fire weather index, as in Canadian Forest Fire Weather Index System (\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e), was calculated at the daily scale from temperature, relative humidity, wind speed and precipitation using \u0026lsquo;cffdrs\u0026rsquo; package(\u003cspan citationid=\"CR70\" class=\"CitationRef\"\u003e70\u003c/span\u003e) in R. The calculation of cumulative water deficit follows Arag\u0026atilde;o et al., 2007 (\u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e68\u003c/span\u003e) using the climate datasets mentioned above.\u003c/p\u003e\u003cp\u003eEight anthropogenic variables, represented by land use and land use change, include forest fraction, pasture fraction, natural non-forest vegetation fraction, cropland fraction, Δforest fraction, Δpasture fraction, Δnatural non-forest vegetation fraction, Δcropland fraction. Fraction refers to the area fraction of each land use type in a grid cell in the current year, and \u0026lsquo;Δfraction\u0026rsquo; is changes in area fraction for a give land use type between two adjacent years.\u003c/p\u003e\u003cp\u003eCollinearity tests of the 15 input features were conducted (Fig. S31).\u003c/p\u003e\u003cp\u003e\u003cb\u003eModel training and validation.\u003c/b\u003e In this study, we first conducted random forest classification to identify grid cells (0.5\u0026deg;\u0026times;0.5\u0026deg;) with non-extreme and extreme fires. We define grid cells with extreme fires as grid cells with burned area exceeding the 90th percentile of the entire time series' observations. The other grid cells with burned area greater than 0 are thus treated as grid cells with non-extreme fires. Extreme fire is usually defined by a percentile threshold of fire radiative power, fire size or fire expansion speed (\u003cspan citationid=\"CR71\" class=\"CitationRef\"\u003e71\u003c/span\u003e, \u003cspan citationid=\"CR72\" class=\"CitationRef\"\u003e72\u003c/span\u003e). As sensitivity tests, we also tried other thresholds (80th, 85th and 95th percentile) to define the extreme fires (Fig. S28, S29).\u003c/p\u003e\u003cp\u003eFor grid cells with non-extreme or extreme fires, we performed random forest quantile regression of burned area separately. The output of quantile regressors is probability distribution, and quantiles of predictions can be retrieved from the distribution. Because the normal random forest algorithm is more conservative, it is not appropriate for predicting burned area of extreme fires due to the small sample number in the training dataset (Fig. S32). Therefore, we used a combination of different quantile predictions to best match the observations. Specifically, two or three output percentiles from 5th, 10th, \u0026hellip;, 95th quantiles were randomly selected as a combination. Then the performance of the selected quantile combination was tested against the observations in the test dataset. Note that 20% of the samples were randomly split for testing, and the remaining 80% samples were used for training. The combination with the highest R\u003csup\u003e2\u003c/sup\u003e and the lowest RMSE was set as the final combination of output quantiles. The R\u003csup\u003e2\u003c/sup\u003e of all quantile combinations during this process are shown in Fig. S24 and S25 for non-extreme fires and in Fig. S26 and S27 for extreme fires.\u003c/p\u003e\u003cp\u003eWe also tried other machine learning methods such as normal random forest regression, gradient boosting regression, extreme gradient boosting regression and long-short-term memory neuron network regression, but their performances (Fig. S33) are not as good as the random forest quantile regression (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eg-j).\u003c/p\u003e\u003cp\u003eIn addition to the 20% samples for testing mentioned above, the model was also validated independently year by year within the time span of burned area observations. For the data in a given year used for validation, data from the remaining years excluding this year were used to train the models, and predictions from the trained models were compared with the observed burned area in this year. Note that a significant positive linear relationship is detected between the R\u003csup\u003e2\u003c/sup\u003e of independent validations and the anomalies of annual burned area (Fig. S34, S35), suggesting our models perform better in years with larger positive burned area anomalies.\u003c/p\u003e\u003cp\u003eAdditionally, we also conducted feature selection with recursive feature elimination cross validation using the RFECV function in \u0026ldquo;scikit-learn\u0026rdquo; package (\u003cspan citationid=\"CR73\" class=\"CitationRef\"\u003e73\u003c/span\u003e) in Python, and the feature selection results are shown in Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e-S3, Fig. S36. The SHapley Additive exPlanations (SHAP) value is calculated with the \u0026ldquo;shap\u0026rdquo; package (\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e) using tree explainer in Python.\u003c/p\u003e\u003cp\u003e\u003cb\u003eCausality tests.\u003c/b\u003e Because land use change (LUC) may cause fires, and fires also result in land cover change, the direction of causality between LUC and burned area variation needs to be further confirmed. To prove whether LUC drives the inter-annual variability of burned area (IAV_BA) rather than the reverse, we applied Convergent Cross Mapping (CCM) (\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e) analysis to test the causality between LUC and IAV_BA. CCM is a causality identification method to distinguish causality from correlation based on non-linear state space reconstruction, which claim to be suitable for non-separable and weakly connected dynamic systems (e.g., ecosystems) (\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e). The CCM coefficient, ranging from \u0026minus;\u0026thinsp;1 to 1, quantifies the strength of causal influence, with larger absolute values indicating stronger causal effects. If the causal effect of A on B (denoted as A:B) is stronger than the reverse (B:A), A is recognized as the cause of B.\u003c/p\u003e\u003cp\u003eThe CCM analysis was conducted on the annual series of ever burned grid cells within the observation period based on the four burned area datasets (FireCCI, MODIS, GABAM and MapBiomas) (Fig. S37). In addition to burned area, area change of forest and pasture, as the two most important land use types according to the SHAP value ranking of machine learning models (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e), were taken into consideration. For each ever-burned grid cell, we extracted annual time series of the selected variables (burned area, ΔForest, and ΔPasture) over the whole observation period, and calculated the CCM coefficients using the \u0026ldquo;CCM\u0026rdquo; function in the \u0026ldquo;pyEDM\u0026rdquo; package (\u003cspan citationid=\"CR74\" class=\"CitationRef\"\u003e74\u003c/span\u003e) in Python. We set no embedding delay and fixed the prediction interval at 1 year, looping the embedding dimension from 1 to 5 to select the optimal parameter. If the CCM coefficient for ΔForest:burned area is larger than that for burned area:ΔForest, then ΔForest is considered to drive burned area change in that grid cell; the same approach was applied to ΔPasture and burned area. The CCM coefficient maps are shown in Fig. S12. When both ΔForest and ΔPasture influence burned area, we selected the larger absolute CCM coefficient (the third column of subplots in Fig. S12).\u003c/p\u003e\u003cp\u003e\u003cb\u003eModel prediction.\u003c/b\u003e After the model training and validation, we used the trained random forest models to predict future burned area during 2021\u0026ndash;2050. Because both climate (CRUJRA and ISIMIP3b) and land use datasets (MapBiomas, LUHv2f, and LuccMEBR) were inconsistent between the historical period (1985\u0026ndash;2020) and the future (2021\u0026ndash;2050), we conducted several sensitivity tests using several harmonization methods to check the robustness of our results (Supplementary Text 1). For the climate datasets, we tested CRUJRA and ISIMIP3b respectively in model training (Supplementary Text 1.1, Fig. S40 and S41). For the land use datasets, land use types in all products were first reclassified to five categories: forest, non-forest natural vegetation, pasture, cropland and others. To harmonize these land use maps from the present day into the future, we used the interannual spatial changes recorded in each future dataset (2020\u0026ndash;2050) and incrementally applied them to the 2020 MapBiomas baseline. Specifically, we calculated the spatial difference maps between two adjacent years from 2020 to 2050 and accumulated them year by year to the 2020 land use map of MapBiomas. In addition to this harmonization method (denoted as \u0026ldquo;spatial difference\u0026rdquo;), we also tested another two harmonization methods (see details in Supplementary Text 1.2). In the harmonization processes, if the harmonized fraction of one land use type exceeds 1 in a grid cell, it remains to be 1 and the fractions of other land use types are set to be 0; if it is less than 0, it is set as 0, and the deficit is allocated to other land use types proportionally to their fractions.\u003c/p\u003e\u003cp\u003eWe performed a set of predictions to separate the contributions of future climate and anthropogenic activities to burned area under the SSP126 and SSP585, including predictions using 1) fixed climate (the average state during 2011\u0026thinsp;~\u0026thinsp;2020) and land use (the same as 2020), 2) fixed climate but dynamic land use changes, 3) fixed land use but transient climate conditions and 4) transient climate conditions and dynamic land use changes. The anthropogenic effect is calculated from the difference between the first two scenarios, and climate effect is from the difference between the first and the third scenarios. The interaction between climate change and anthropogenic activities are calculated as the difference between the first and the fourth scenarios minus the climate effect and anthropogenic effect.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAcknowledgements\u003c/h2\u003e\u003cp\u003eThis study was supported by Yunnan Provincial Science and Technology Project at Southwest United Graduate School (grant number: 202302AO370001), the National Natural Science Foundation of China (grant number: 42175169), the Hainan Institute of National Park Research Program KY-23ZK01, and the CALIPSO (Carbon Loss In Plants, Soils and Oceans), funded through Schmidt Sciences LLC. SS, SB, PC were supported by ESA XFires (contract number 4000145351/24/I-LR).\u003c/p\u003e\n\u003ch3\u003eData Availability\u003c/h3\u003e\n\u003cp\u003eFireCCI51 can be downloaded from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://geogra.uah.es/fire_cci/firecci51.php\u003c/span\u003e\u003cspan address=\"https://geogra.uah.es/fire_cci/firecci51.php\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. MCD64CMQ can be downloaded from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://lpdaac.usgs.gov/products/mcd64a1v006/\u003c/span\u003e\u003cspan address=\"https://lpdaac.usgs.gov/products/mcd64a1v006/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. MapBiomas fire can be downloaded from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://brasil.mapbiomas.org/en/mapbiomas-fogo/\u003c/span\u003e\u003cspan address=\"https://brasil.mapbiomas.org/en/mapbiomas-fogo/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Global Annual Burned Area Map (GABAM) can be downloaded from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://vapd.gitlab.io/post/gabam/\u003c/span\u003e\u003cspan address=\"https://vapd.gitlab.io/post/gabam/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. MapBiomas land use maps can be downloaded from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://brasil.mapbiomas.org/en/colecoes-mapbiomas/\u003c/span\u003e\u003cspan address=\"https://brasil.mapbiomas.org/en/colecoes-mapbiomas/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Land-Use Harmonization version 2-future (LUH v2f) datasets can be downloaded from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://luh.umd.edu/data.shtml\u003c/span\u003e\u003cspan address=\"https://luh.umd.edu/data.shtml\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. LuccME/INLAND land-use scenarios for Brazil 2050 can be downloaded from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://zenodo.org/records/10611737\u003c/span\u003e\u003cspan address=\"https://zenodo.org/records/10611737\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. CRUJRA v2.2 can be downloaded from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://catalogue.ceda.ac.uk/uuid/4bdf41fc10af4caaa489b14745c665a6\u003c/span\u003e\u003cspan address=\"https://catalogue.ceda.ac.uk/uuid/4bdf41fc10af4caaa489b14745c665a6\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. ISIMIP3b IPSL output can be downloaded from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.isimip.org/gettingstarted/data-access/\u003c/span\u003e\u003cspan address=\"https://www.isimip.org/gettingstarted/data-access/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eLiu YY et al (2015) Recent reversal in loss of global terrestrial biomass. Nat Clim Change 5:470\u0026ndash;474\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMalhi Y et al (2008) Climate Change, Deforestation, and the Fate of the Amazon. Science 319:169\u0026ndash;172\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBarlow J et al (2016) Anthropogenic disturbance in tropical forests can double biodiversity loss from deforestation. Nature 535:144\u0026ndash;147\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSkole D, Tucker C (1993) Tropical Deforestation and Habitat Fragmentation in the Amazon: Satellite Data from 1978 to 1988. Science 260:1905\u0026ndash;1910\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBarlow J, Berenguer E, Carmenta R, Fran\u0026ccedil;a F (2020) Clarifying Amazonia's burning crisis. Glob Change Biol 26:319\u0026ndash;321\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBond WJ, Woodward FI, Midgley GF (2005) The global distribution of ecosystems in a world without fire. New Phytol 165:525\u0026ndash;538\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLasslop G, Brovkin V, Reick CH, Bathiany S, Kloster S (2016) Multiple stable states of tree cover in a global land surface model due to a fire-vegetation feedback. Geophys Res Lett 43:6324\u0026ndash;6331\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRanderson JT et al (2006) The Impact of Boreal Forest Fire on Climate Warming. Science 314:1130\u0026ndash;1132\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003evan der Werf GR et al (2010) Global fire emissions and the contribution of deforestation, savanna, forest, agricultural, and peat fires (1997\u0026ndash;2009). Atmos Chem Phys 10:11707\u0026ndash;11735\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eArag\u0026atilde;o LEOC, Shimabukuro YE (2010) The Incidence of Fire in Amazonian Forests with Implications for REDD. Science 328:1275\u0026ndash;1278\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMataveli G et al (2022) Record-breaking fires in the Brazilian Amazon associated with uncontrolled deforestation. Nat Ecol Evol 6:1792\u0026ndash;1793\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePivello VR (2011) The Use of Fire in the Cerrado and Amazonian Rainforests of Brazil: Past and Present. Fire Ecol 7:24\u0026ndash;39\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePendrill F et al (2022) Disentangling the numbers behind agriculture-driven tropical deforestation. Science 377:eabm9267\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAndela N et al Tracking and classifying Amazon fire events in near real time. Sci Adv 8, eabd2713\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eda Silva Junior CA et al (2020) Persistent fire foci in all biomes undermine the Paris Agreement in Brazil. Sci Rep 10:16246\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen Y et al (2011) Forecasting Fire Season Severity in South America Using Sea Surface Temperature Anomalies. Science 334:787\u0026ndash;791\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFonseca MG et al (2017) Climatic and anthropogenic drivers of northern Amazon fires during the 2015\u0026ndash;2016 El Ni\u0026ntilde;o event. Ecol Appl 27:2514\u0026ndash;2527\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eArag\u0026atilde;o LEOC et al (2018) 21st Century drought-related fires counteract the decline of Amazon deforestation carbon emissions. Nat Commun 9:536\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCai W et al (2020) Climate impacts of the El Ni\u0026ntilde;o\u0026ndash;Southern Oscillation on South America. Nat Reviews Earth Environ 1:215\u0026ndash;231\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang Y, Huang P (2022) Potential fire risks in South America under anthropogenic forcing hidden by the Atlantic Multidecadal Oscillation. Nat Commun 13:2437\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNepstad DC et al (1999) Large-scale impoverishment of Amazonian forests by logging and fire. Nature 398:505\u0026ndash;508\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFearnside PM (2005) Deforestation in Brazilian Amazonia: History, Rates, and Consequences. Conserv Biol 19:680\u0026ndash;688\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBalch JK, Nepstad DC, Curran LM (2009) In: Cochrane MA (ed) Tropical Fire Ecology: Climate Change, Land Use, and Ecosystem Dynamics. Springer Berlin Heidelberg, Berlin, Heidelberg, pp 481\u0026ndash;502\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBowring SPK, Li W, Mouillot F, Rosan TM, Ciais P (2024) Road fragment edges enhance wildfire incidence and intensity, while suppressing global burned area. Nat Commun 15:9176\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSills J, Arruda D, Candido HG, Fonseca R (2019) Amazon fires threaten Brazil's agribusiness. Science 365:1387\u0026ndash;1387\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBrando PM et al (2014) Abrupt increases in Amazonian tree mortality due to drought\u0026ndash;fire interactions. \u003cem\u003eProceedings of the National Academy of Sciences\u003c/em\u003e 111, 6347\u0026ndash;6352\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eE. Chuviecoet al., Generation and analysis of a new global burned area product based on MODIS 250\u0026thinsp;m reflectance bands and thermal anomalies. \u003cem\u003eEarth Syst. Sci. Data\u003c/em\u003e 10, 2015\u0026ndash;2031 (2018)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGiglio L, Boschetti L, Roy DP, Humber ML, Justice CO (2018) The Collection 6 MODIS burned area mapping algorithm and product. Remote Sens Environ 217:72\u0026ndash;85\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMapBiomas-Project (2021) Fire Scars (Collection 2 - MapBiomas Fire). (MapBiomas Project\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLong T et al (2019) 30 m Resolution Global Annual Burned Area Mapping Based on Landsat Images and Google Earth Engine. Remote Sens 11:489\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGuo Z et al (2025) Reconstructed global monthly burned area maps from 1901 to 2020. \u003cem\u003eEarth Syst. Sci. Data Discuss.\u003c/em\u003e 1\u0026ndash;28 (2025)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eO\u0026rsquo;Neill BC et al (2014) A new scenario framework for climate change research: the concept of shared socioeconomic pathways. Clim Change 122:387\u0026ndash;400\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003evan Vuuren DP et al (2011) The representative concentration pathways: an overview. Clim Change 109:5\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSugihara G et al (2012) Detecting Causality in Complex Ecosystems. Science 338:496\u0026ndash;500\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLima CHR, AghaKouchak A, Randerson JT (2018) Unraveling the Role of Temperature and Rainfall on Active Fires in the Brazilian Amazon Using a Nonlinear Poisson Model. J Geophys Research: Biogeosciences 123:117\u0026ndash;128\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJain P, Castellanos-Acuna D, Coogan SCP, Abatzoglou JT, Flannigan MD (2022) Observed increases in extreme fire weather driven by atmospheric humidity and temperature. Nat Clim Change 12:63\u0026ndash;70\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBalch JK et al (2022) Warming weakens the night-time barrier to global fire. Nature 602:442\u0026ndash;448\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNolan RH et al (2016) Predicting dead fine fuel moisture at regional scales using vapour pressure deficit from MODIS and gridded weather data. Remote Sens Environ 174:100\u0026ndash;108\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGrossiord C et al (2020) Plant responses to rising vapor pressure deficit. New Phytol 226:1550\u0026ndash;1566\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLundberg SM et al (2020) From local explanations to global understanding with explainable AI for trees. Nat Mach Intell 2:56\u0026ndash;67\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e\u003cem\u003eEquations and FORTRAN program for the Canadian Forest Fire Weather Index System\u003c/em\u003e (1985) \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://cfs.nrcan.gc.ca/publications?id=19973\u003c/span\u003e\u003cspan address=\"https://cfs.nrcan.gc.ca/publications?id=19973\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMaeda EE et al (2021) Large-scale commodity agriculture exacerbates the climatic impacts of Amazonian deforestation. \u003cem\u003eProceedings of the National Academy of Sciences\u003c/em\u003e 118, e2023787118\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDunne JP et al (2020) The GFDL Earth System Model Version 4.1 (GFDL-ESM 4.1): Overall Coupled Model Description and Simulation Characteristics. \u003cem\u003eJournal of Advances in Modeling Earth Systems\u003c/em\u003e 12, e2019MS002015\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBoucher O et al (2020) Presentation and Evaluation of the IPSL-CM6A-LR Climate Model. J Adv Model Earth Syst 12:e2019MS002010\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGutjahr O et al (2019) Max Planck Institute Earth System Model (MPI-ESM1.2) for the High-Resolution Model Intercomparison Project (HighResMIP). Geosci Model Dev 12:3241\u0026ndash;3281\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYukimoto S, The Meteorological Research Institute Earth System Model Version et al (2019) 2.0, MRI-ESM2.0: Description and Basic Evaluation of the Physical Component. J Meteorological Soc Japan Ser II 97:931\u0026ndash;965\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSellar AA et al (2019) UKESM1: Description and Evaluation of the U.K. Earth System Model. J Adv Model Earth Syst 11:4513\u0026ndash;4558\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLange S, B\u0026uuml;chner M (2021) ISIMIP3b bias-adjusted atmospheric climate input data (v1.1). ISIMIP Repository\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHurtt GC et al (2020) Harmonization of global land use change and management for the period 850\u0026ndash;2100 (LUH2) for CMIP6. Geosci Model Dev 13:5425\u0026ndash;5464\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSilva Bezerra FG et al (2022) New land-use change scenarios for Brazil: Refining global SSPs with a regional spatially-explicit allocation model. PLoS ONE 17:e0256052\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCochrane MA (2003) Fire science for rainforests. Nature 421:913\u0026ndash;919\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTyukavina A et al (2017) Types and rates of forest disturbance in Brazilian Legal Amazon, 2000\u0026ndash;2013. Sci Adv 3:e1601047\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSilv\u0026eacute;rio DV et al (2013) Testing the Amazon savannization hypothesis: fire effects on invasion of a neotropical forest by native cerrado and exotic pasture grasses. Philosophical Trans Royal Soc B: Biol Sci 368:20120427\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCornwall W (2022) Fire Invasions. Around the world, flammable invasive grasses are increasing the risks of damaging wildfires. Science 377\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHaddad NM et al (2015) Habitat fragmentation and its lasting impact on Earth\u0026rsquo;s ecosystems. Sci Adv 1:e1500052\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBrinck K et al (2017) High resolution analysis of tropical forest fragmentation and its impact on the global carbon cycle. Nat Commun 8:14855\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBrando PM, Oliveria-Santos C, Rocha W, Cury R, Coe MT (2016) Effects of experimental fuel additions on fire intensity and severity: unexpected carbon resilience of a neotropical forest. Glob Change Biol 22:2516\u0026ndash;2525\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRosan TM et al (2022) Fragmentation-Driven Divergent Trends in Burned Area in Amazonia and Cerrado. Front Forests Global Change 5\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhao Z et al (2021) Fire enhances forest degradation within forest edge zones in Africa. Nat Geosci 14:479\u0026ndash;483\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCochrane MA (2001) Synergistic Interactions between Habitat Fragmentation and Fire in Evergreen Tropical Forests. Conserv Biol 15:1515\u0026ndash;1521\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eArmenteras D, Gonz\u0026aacute;lez TM, Retana J (2013) Forest fragmentation and edge influence on fire occurrence and intensity under different management types in Amazon forests. Biol Conserv 159:73\u0026ndash;79\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHarris I, Jones PD, Osborn TJ, Lister DH (2014) Updated high-resolution grids of monthly climatic observations \u0026ndash; the CRU TS3.10 Dataset. Int J Climatol 34:623\u0026ndash;642\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKobayashi S et al (2015) The JRA-55 Reanalysis: General Specifications and Basic Characteristics. J Meteorological Soc Japan Ser II 93:5\u0026ndash;48\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHarris I, Osborn TJ, Jones P, Lister D (2020) Version 4 of the CRU TS monthly high-resolution gridded multivariate climate dataset. Sci Data 7:109\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMapBiomas-Project (2022) Collection 6 of the Annual Series of Land Use and Land Cover Maps of Brazil. (MapBiomas Project\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCucchi M et al (2020) WFDE5: bias-adjusted ERA5 reanalysis data for impact studies. Earth Syst Sci Data 12:2097\u0026ndash;2120\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAguiar AP, Soler AP, Assis L (2011) T., in \u003cem\u003eGLP Land Use Transitions in South America: framing the present, preparing for a sustainable future Workshop\u003c/em\u003eIlhabela\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eArag\u0026atilde;o LEOC et al (2007) Spatial patterns and fire response of recent Amazonian droughts. Geophys Res Lett 34\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTetens O (1930) Uber einige meteorologische begriffe. Z Geophys 6\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang X et al (2017) cffdrs: an R package for the Canadian Forest Fire Danger Rating System. Ecol Processes 6:5\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBowman DMJS et al (2017) Human exposure and sensitivity to globally extreme wildfire events. Nat Ecol Evol 1:0058\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFrancisco Castro PM, Rego P, Fernandes C, Hoffman (2021) \u003cem\u003eFire Science\u003c/em\u003e. Springer Textbooks in Earth Sciences, Geography and EnvironmentSpringer Cham, ed. 1, pp. XXXVIII, 644\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePedregosa F et al (2011) Scikit-learn: Machine Learning in Python. J Mach Learn Res 12:2825\u0026ndash;2830\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePark J, GitHub (2019) \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/SugiharaLab/pyEDM\u003c/span\u003e\u003cspan address=\"https://github.com/SugiharaLab/pyEDM\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"nature-portfolio","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"","title":"Nature Portfolio","twitterHandle":"","acdcEnabled":false,"dfaEnabled":false,"editorialSystem":"ejp","reportingPortfolio":"","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-7604250/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7604250/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe Brazilian Amazon suffers from frequent fires with varying sizes and intensities, but the causes for extreme and non-extreme fires remain unclear. Here, using multiple satellite-based observations and explainable machine learning models, we find distinct drivers of burned area variation for extreme (burned area\u0026thinsp;\u0026gt;\u0026thinsp;the 90th percentile) and non-extreme fires in the Brazilian Amazon between 1985 and 2020. The absolute land use fraction of pasture and forest are dominant drivers for non-extreme fires, while the extreme fires are more driven by the fire weather conditions and land use change from forest to pasture in the adjacent two years. In future climate and land use change scenarios, our predicted annual total burned area from extreme fires and non-extreme fires increases from 2021 to 2050. Compared with the historical period, contributions of future climate change and anthropogenic activities to the annual total burned area are positive for non-extreme fires but negative for extreme fires due to reduced pasture expansion and deforestation. Therefore, mitigating climate change and implementing local sustainable land use strategies are crucial for restricting fires in the Brazilian Amazon.\u003c/p\u003e","manuscriptTitle":"Distinct drivers of extreme and non-extreme fires in the Brazilian Amazon","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-29 06:45:18","doi":"10.21203/rs.3.rs-7604250/v1","editorialEvents":[],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"nature-communications","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"NCOMMS","sideBox":"Learn more about [Nature Communications](http://www.nature.com/ncomms/)","snPcode":"","submissionUrl":"https://mts-ncomms.nature.com/","title":"Nature Communications","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Nature Communications","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"0e011be7-f615-4ff5-9e27-5526872df57c","owner":[],"postedDate":"September 29th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":55387420,"name":"Earth and environmental sciences/Ecology/Fire ecology"},{"id":55387421,"name":"Earth and environmental sciences/Ecology/Ecosystem ecology"}],"tags":[],"updatedAt":"2025-11-21T11:06:18+00:00","versionOfRecord":[],"versionCreatedAt":"2025-09-29 06:45:18","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7604250","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7604250","identity":"rs-7604250","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}