Changes in land use enhance the sensitivity of tropical ecosystems to fire-climate extremes

The Pantanal, the largest contiguous wetland in the world with a high diversity of ecosystems and habitat for several endangered species, was impacted by record-breaking wildfires in 2020. In this study, we integrate satellite and modeling data that enable exploration of natural and human contributing factors to the unprecedented 2020 fires. We demonstrate that the fires were fueled by an exceptional multi-year drought, but dry conditions solely could not explain the spatial patterns of burning. Our analysis reveals how human-caused fires exacerbated drought effects on natural ecosystem within the Pantanal, with large burned fractions primarily over natural (52%), and low cattle density areas (44%) in 2020. The post-fire ecosystem and hydrology changes also had strong ecological effects, with vegetation productivity less than − 1.5 σ over more than 30% of the natural and conservation areas. In contrast to more managed areas, there was a clear decrease in evaporation (by ~ 9%) and an increase in runoff (by ~ 5%) over the natural areas, with long-term impacts on ecosystem recovery and fire risk. This study provides the first tropical evidence outside rainforests of the synergy between climate, land management and fires, and the associated impacts on the ecosystem and hydrology over the largest contiguous wetlands in the world.

The driving factors behind the 2020 Pantanal fires. Over 71% of the region, about 9.2 Mha, face standardized root zone soil moisture anomaly values below − 1 σ , indicating the unprecedented nature of the 2020 drought, both in terms of the magnitude and spatial extent (Fig. 1d). We examine the severity and extent of the recent drought over the Pantanal using deficits in antecedent precipitation and root zone soil moisture as analogs of drought conditions. Precipitation shortages represent the meteorological factors driving the anomalous dry conditions. Deficits of soil moisture resulting from the imbalance between moisture supply on the land surface and the losses from evaporation and runoff are typically used to characterize agricultural droughts [48][49][50] . Maps of root zone soil moisture anomalies for the dry season (Aug-Nov) across 2017 to 2020 (Fig. S1) indicate that, prior to 2018, the small root zone soil moisture anomalies are representative of the typical dry season. The drought onset is noticed in 2018, and by 2019 much of the region is in consistent negative soil moisture anomalies, exacerbating in 2020.
The sensitivity of protected and extensively used land to the 2020 extreme fire occurrence is further highlighted by time series analysis. Figure 2f shows that the correlation between antecedent precipitation and burned area range from − 0.16 to 0.12 and the root zone soil moisture and burned area range from − 0.21 to − 0.03 during 2003-2019. Over the very low pasture areas, the association between precipitation and burned areas increases significantly (from − 0.16 to − 0.24) when 2020 data are considered, whereas no considerable change in the medium and high pasture density areas is seen. These results highlight the extreme and unequal impact of the 2020 drought, but also suggest that the sensitivity of natural ecosystems to fire has increased over time. Indeed, the observed trends in fraction of burned areas in the very low, low, and medium pasture areas were not significant during 2003-2019 (Table 1), but when 2020 data are included, there is a statistically significant increasing trend in the fraction of burned area over the protected and very low-density pasture areas. In contrast, no change is observed over the low and medium density areas where the relationship between burned area and drought remains weak. It is also notable that a statistically significant decreasing trend in the fire occurrences is seen over the high-density pasture areas, which is unaffected by the inclusion or exclusion of the 2020 data, consistent with increasing industrialization and cattle densities over time.
Similar results are obtained with comparisons of the fraction of burned areas and the associated root zone soil moisture anomalies over the four main land cover types (Fig. S2). The forests and wetlands are mostly natural areas whereas the crop, grass, and savanna areas are, to a large part, used for grazing. The natural vegetation areas show significantly higher percent of burns in 2020 compared to such occurrences in the observational www.nature.com/scientificreports/ larger than the fire occurrences in the MCD64A1 data record. In the years prior to 2020, the percent of burns in the conserved areas never exceeded 18%. Patterns similar to Figure 2f are observed when the level of association between antecedent moisture conditions and the fraction of burned area is stratified by the main land cover types and over the conservation and non-conservation areas (Fig. S3). Compared to the 2003-2019 period, there is a statistically significant change in the correlation between antecedent precipitation with the percent of burned area over the forests (a decrease of 36%) and wetlands (a decrease of 52%) when data from year 2020 is included (Fig. S3a). On the other hand, over  www.nature.com/scientificreports/ the crop, grass and savanna areas, the changes in correlation between the antecedent precipitation anomalies and burned area with and without 2020 are not statistically significant. Similar patterns are observed between areas classified as reserves and non-reserves (Fig. S3b). Over the reserves, there is a statistically significant increase in the association between antecedent precipitation and the fraction of burned areas, when the 2020 record is considered, whereas over non-reserves no statistically significant change in the fire activity is observed for 2020. We also examined the influence of large climate patterns (Fig. S3c) by examining the changes in correlation between the percentage of burned areas with the climate indices of the Atlantic Multidecadal Oscillation (AMO 51 ), the Pacific Decadal Oscillation (PDO 52 ) and the Multivariate El Nino Index (MEI 53 ). The AMO represents the anomalous warming of the Atlantic Ocean, whereas PDO and MEI represent the anomalous warming of the equatorial and eastern tropical Pacific and tropical north Atlantic oceans, respectively. Consistent with prior studies over Amazon 31,54 and Pantanal 40 , there is a general association between these climate indices and the drought and fire instances in the Pantanal. Though these correlation values show a marginal change when the year 2020 is included, compared to the period of 2002-2019, these changes are not statistically significant, suggesting that the increase in fire activity in 2020 is not solely explained by the anomalies in large scale climate variability indices. In contrast to the widely held assumption that remote areas are well protected from ecosystem degradation, our results highlight how the natural environments are more prone to fires under extreme drought conditions and human-induced ignitions, such as in 2020. Under such scenarios, fires spread over large areas without any suppression, increasing the difficulty of controlling fires 42 . Conservation strategies such as domesticated cattle exclosures will be less effective in such areas, under scenarios of future climate change with more frequent occurrences of dry extremes 55 .
To confirm the inferences from the correlation analysis, a Random Forest (RF) classifier model using the pasture density, landcover type, standardized anomalies of precipitation, root zone soil moisture, Leaf Area Index (LAI), and Gross Primary Production (GPP) as predictors and the burned area locations as the predictand is used (the larger scale climate indicators are not included given their lack of specific influence in the 2020 fires). The ordering of the importance of the predictors is generally consistent (Fig. S4) with the results from the onevariable at a time regression estimates described above, with the pasture density and the antecedent root zone soil moisture anomalies being the predictors with larger importance values. For example, the drop column importance values for the pasture density and antecedent root zone soil moisture are 0.28 each, higher than 0.21 for precipitation, 0.18 for landcover, 0.04 for LAI and 0.0 for GPP. The relative importance of the pasture density from this analysis also points to the role of human management in the fires over the Pantanal. This suggests human-caused fires often escape beyond their intended area in natural landscapes while controlled fires over areas with more grazing and pasture are not necessarily influenced by dry and drought conditions. Ecosystem and hydrology impacts from the 2020 Pantanal fires. The impact of fires on the ecosystem and hydrological conditions had strong ecological effects, though the exceptional drought is largely uniform across the whole region. Except for 2020, the correspondence between drought and negative vegetation anomalies is weak in this region ( Fig. 3a; S5a-b), indicating that the Pantanal ecosystem is very resilient to moisture stress conditions. Moreover, while the entire region is in drought, the significant drop in LAI and GPP is primarily observed over areas such as the forests, wetlands, and conservation areas (Fig. 3b), with close correspondence to burned areas during 2020. Over the non-forested areas, a more moderate drop in LAI and GPP is observed. The results of this study also show that changes in precipitation are not the main driver of vegetation variability, consistent with prior studies that indicated that rainfall is not a good predictor of local vegetation structure in areas that flood seasonally 56 . The correlation between the standardized LAI and GPP anomalies and antecedent precipitation anomalies ( Fig. S5a-b) is generally small (< 0.10), particularly in the 2002-2019 period. The significant negative anomalies in precipitation and vegetation states in 2020 increase these associations when the entire period is considered. This pattern is observed in the stratification with the dominant vegetation types and the pasture areas.
Our results show that antecedent vegetation conditions are not a major factor in the development of fires, rather, negative vegetation anomalies are observed following the fires (Fig. S5c-d). Compared to precipitation and root zone soil moisture, the antecedent vegetation anomalies are more weakly correlated with the burn areas. There is no statistically significant relationship between the antecedent LAI and burned areas prior to 2020, whereas the correlation between antecedent GPP and burned areas show little change with and without 2020, both over different land cover and pasture regimes. This is an indication that the potential fuel loads from the negative vegetation anomalies are not a major factor in the fire occurrences. On the other hand, the 2-month leading LAI and GPP anomalies show stronger correlations with the burned areas, indicating that the negative anomalies in vegetation variables follow the fire related burns or are consequences of the fires. The changes in the correlation between leading LAI/GPP and burn area in the 2002-2019 and 2002-2020 are statistically significant over the forests, confirming that the vegetation disturbances follow the fires over these areas. Over the crop, grassland, and savannas, the change in the correlations with and without 2020 is not statistically significant. During the 2003-2019 period, the average % of conservation areas with standardized LAI (GPP) values below − 1.5 σ is only 3.3 (1.5), whereas in 2020, the % of areas with such large vegetation alterations jump to 35.6 (28.2) ( Table 2). Similar large changes in vegetation anomalies are seen in areas with historically very low pasture density and forests, whereas over other areas, more moderate reductions in vegetation productivity are seen.
The annual average precipitation in 2020 was 40% below average fueling the drought and wildfires. A number of precipitation events in late October and early November, 2020 eventually helped in extinguishing the fires. During those rain events, the changes in ET and runoff had strong ecological effects with increasing runoff (with the interquartile range of 3-9%) and reduced ET (with the interquartile range of 0 to − 15%) over the burned areas of forests, conservation, and very low-density pasture (Fig. 3c), confirming that the removal of vegetation www.nature.com/scientificreports/ due to fire impacted the local hydrological response. Over other areas, the magnitude of such changes is smaller with a net impact of no clear shift in the hydrological regime (Fig. S6). Note that the increasing runoff and reducing ET impact from fires quantified here is likely underestimated, as factors such as water repellence of the soil from fires 57 , are not represented in our modeling system. As a consequence of the increased runoff, increased erosion and sediment yield can further contribute to the land degradation. The drastic changes in the local hydrology from the 2020 fires in the Pantanal underscore how the combination of droughts and human caused  www.nature.com/scientificreports/ fires can have long lasting effects on the hydrological cycle and regional climate. These changes in turn form part of a positive feedback loop, with drier and warmer conditions, more open cover types, and enhanced fire risk.

Implications.
Understanding the complex feedback mechanisms involving fire, climate, vegetation, and human activity is crucial, but can only be achieved when there is comprehensive knowledge about these patterns. The sensitivity of natural landscapes to fire-driven degradation has been a concern across the southern Amazon for years, but here we demonstrate that the same mechanisms may be more universally applicable across the tropics. This study provides the first tropical evidence outside rainforests of the synergy between climate, land management and fires, and the associated impacts on the ecosystem and hydrology over the largest contiguous wetlands in the world. Though similar levels of drought anomalies have been observed in the past decades, the temporal extent of the 2020 drought exceeds what has been observed in prior years. Nevertheless, climate alone could not explain the unprecedented extent and location of wildfires in the Pantanal, in particularly during 2020, as previous observed over the Amazon rainforest 16,31 . Here, we demonstrate how cattle density, and associated shifts in land use drive regional fire patterns and its sensitivity to climate. Strikingly, natural areas, not humandominated landscapes, were most sensitive to fire driven ecosystem degradation. While such conditions have been predicted to occur across parts of the southern Amazon 34,35,58-61 , here we provide large-scale evidence from the observational record. In addition, the biome-scale extent of the fires may further slow ecosystem recovery through its impacts on regional hydrology and climate. We observed a 30% decrease in the vegetation productivity, 9% decrease in ET and a 5% increase in runoff across the burned regions presenting an important feedback mechanism 12,62,63 . Our findings have global implications, as they demonstrate how tropical land management and fires impact the regional water cycle. Combined, these two factors can result in rapid degradation of natural areas previously assumed to be protected by their remoteness. Based on these findings, there is an urgent need to rethink tropical ecosystem conservation and adaptation strategies, using a framework that acknowledges the compounding, cascading and long-term effects of land management and climate. The quantification of such long-term impact assessments is left for a future work. The input meteorological fields of air temperature, humidity, surface pressure, wind, downward shortwave radiation, and downward longwave radiation are downscaled to 1-km by adjusting for terrain differences in elevation, slope, and aspect 70 . Precipitation fields at 1-km are generated by using the monthly high resolution WorldClim climatology 71 74 . In this study, we assimilate the 500m MCD15A2H collection 6 leaf area index (LAI) retrievals from the Moderate Resolution Imaging Spectroradiometer (MODIS) sensors aboard NASA's Terra and Aqua satellites, using this feature in Noah-MP. Similar to the studies that describe the assimilation of MODIS LAI data 75,76 , a 1-dimensional Ensemble Kalman Filter (EnKF) algorithm is used to assimilate the LAI retrievals, with a 20 member model ensemble. The ensemble spread is created by applying small perturbations to the model states and meteorological inputs 75 . During assimilation, the prognostic leaf mass variable in the LSM is updated in response to the LAI inputs from the MODIS data. Before assimilation, the MCD15A2H data is aggregated to 1-km resolution and only data values flagged as 'good quality' considering factors such as cloud contamination, algorithm saturation, and detector signal quality are used in assimilation.

Modeling and data assimilation setup.
To evaluate the impact of the vegetation disturbances from the 2020 Pantanal fires on the local hydrology, an additional integration is conducted where the daily climatological vegetation conditions across the 2002-2019 time period are assimilated into the Noah-MP model (called the LAI-climo DA) during the year 2020. As the difference between the LAI-DA and LAI-climo-DA model runs is solely the vegetation changes from the 2020 fires, these two integrations are used to quantify the influence of the fire driven vegetation changes.
A multi-variable random forest (RF) classifier model 77 focusing on the local scale variables, was developed to consider the joint or cross influence of the human management, fuel moisture and fuel load variables to fire occurrences over the Pantanal. RF models are non-parametric and are shown to work well with correlated and conditional variables and are robust in the presence of outliers and noise in the data [78][79][80] . RF also generates measures of feature importance, a score to indicate the usefulness of the input variable at predicting the target. www.nature.com/scientificreports/ Three different approaches are used to compute the feature importances generated by the RF model. The native approach with RF models computes the feature importances based on how much each feature decreases the variances of the optimality criteria used in the model. Similarly, the permutation feature importance method computes the feature importances by randomly reshuffling while preserving the distribution of each predictor to assess its influence on the model performance. The third approach, drop-column importance, excludes each feature one at a time and examines how it impacts the predictive accuracy. The characterization of the extremes in hydrology and ecosystem variables is captured based on the outputs from the model integrations assimilating remotely sensed LAI. Deficits in precipitation and root zone soil moisture from the model integrations are assumed to represent the meteorological and agricultural droughts 48,81 . Specifically, the 12-month standard anomalies in precipitation and 6-month standardized anomalies in the top 1m root zone soil moisture across 2003-2020 are used as analogs of drought. This 12 month precipitation lag window is chosen based on the convention used at the NOAA National Centers for Environmental Information to characterize long-term and persistent droughts. A six-month and two-month window is chosen for capturing agricultural droughts based on the assumed temporal limits soil moisture and vegetation memory [82][83][84] , respectively.
Ancillary datasets. A number of ancillary datasets are used in the analysis and evaluation of the results presented in the article. A map representing the variation in the potential density of cattle head per ranch (certified in the Cadastro Ambiental Rural in the Brazilian Pantanal) 37 is used as an index of the level of pasture in this area. This map is estimated by the Mapbiomas project 85 (https:// mapbi omas. org/), which is a multi-institutional initiative to map the land use dynamics in Brazil and other tropical countries based on information from remote sensing and local information. There are large ranches in this region with 62.3% or more than 5000 ha 37 . As shown in Fig. 1, we define four categories of pasture with cattle heads per hectare ranging from 0 to 0.03, 0.03 to 0.1, 0.1 to 0.4, and > 0.4 for very low, low, medium, and high pasture density areas, respectively.
The impact of assimilating LAI data on various water and carbon fluxes and states is evaluated by comparing the model simulations with a number of remote sensing-based global reference data products (Extended Analysis 1). Soil moisture states from the model are evaluated by comparing against the Level 2 retrievals from NASA's Soil Moisture Active Passive (SMAP 86 ) mission. In addition to soil moisture, SMAP also provides estimates of vegetation optical depth (VOD), an analog of the above ground canopy biomass 87 , derived as part of the radiometric soil moisture retrieval. Here we use both VOD and soil moisture from the SMAP SPL2SMP_E product, which is available at 9-km spatial resolution through Backus-Gilbert interpolation applied to oversampled antenna measurements 88 . To evaluate the evapotranspiration (ET) estimates, MODIS thermal infrared (TIR) data based ET from the Atmosphere-Land Exchange Inverse (ALEXI 89,90 ) model is used. The 5-km resolution gridded ALEXI ET estimates are used in this study. The subsurface moisture simulations are evaluated by comparing the monthly terrestrial water storage anomalies against those from NASA's Gravity Recovery and Climate Experiment (GRACE) and GRACE-Follow On (GRACE-FO) satellites. Specifically, we use the Release 06 GRACE 91 Level-2 Mass Concentration blocks (mascons) data products available from the Center for Space Research (CSR) at University of Texas. This product is available at 0.25 deg spatial resolution at monthly intervals. The impact of LAI DA on carbon fluxes is evaluated by comparing the simulated Gross Primary Production (GPP), which represents the total carbon fixation by plants through photosynthesis, against two remote sensing-based estimates: (1) remote sensing retrievals of Solar Induced Fluorescence (SIF) from the Global Ozone Monitoring Experiment-2 (GOME-2) aboard the MetOp-A satellite 92,93 and (2) MODIS reflectance-based FLUXSAT product 94 . SIF represents the amount of solar radiation absorbed by chlorophyll and reemitted as fluorescence and is considered as a functional analog of GPP. FLUXSAT is developed through a data driven modeling approach by calibrating the MODIS reflectance information to ground measurements from the FLUXNET network. For uniformity of comparison, all reference data products are interpolated to the 1-km modeling grid in these evaluations.