Main

Tropical dry forests, woodlands and savannahs (hereafter, tropical dry woodlands) cover 20% of the terrestrial biosphere1, harbour extraordinary biodiversity and are major carbon stocks2. Fires are widespread in these systems owing to the abundance of fuel during the dry season and frequent natural ignitions, most importantly lightning3. Consequently, fire is a key ecological process shaping tropical dry woodlands4, including their vegetation structure3, nutrient recycling2, net primary production5 and phenology6. Furthermore, fires in tropical dry woodlands impact the global biosphere, contributing nearly 65% of all fire-related greenhouse gas emissions7. Understanding the fire dynamics of these systems is therefore important.

Although fires occur in all tropical dry woodlands, their extent and frequency vary naturally from region to region8. One reason is variable primary productivity and fuel loads—for instance, due to geographic variations in rainfall or nutrients9. A second major source of variability is weather, particularly drought frequency and duration10,11. Moreover, there are regionally varying interactions between vegetation, fire and herbivory that influence fire regimes in major ways—for instance, whether herbivores reduce fuel loads or whether a lack of herbivory leads to fuel accumulation12. These factors interact, creating complex spatiotemporal fire patterns.

People and their activities add further complexity and can majorly alter fire dynamics. People have used fire for millennia to keep dry woodland landscapes open for agriculture, grazing or hunting13,14. Likewise, people have massively reduced the impact of natural herbivory through the selective hunting and extirpation of native megaherbivores from many dry woodlands, potentially contributing to higher fuel loads and intensifying fires15,16. The expansion of modern agriculture has recently altered fire regimes in tropical dry woodlands17,18. Here, fire is used to clear land, providing a cheap method to remove residual woody biomass after valuable trees have been logged19. Fire is also widely used as an inexpensive management tool20—for example, to remove encroaching shrubs and revitalize soils on pastures21 or to remove crop residues before the next cropping rotation22,23. In contrast, land use in tropical dry woodlands can also lead to fire suppression—for example, because livestock grazing or browsing reduces fuel loads24 or because land managers actively fight fires. Although fire patterns and drivers have been studied extensively22, the impact of land management on fire remains incompletely understood, particularly at broader spatial scales25.

Globally, fire activity has declined over the past few decades26, but it has increased regionally in many tropical dry woodlands12,27. However, whether this trend is due to changes in climate or land use remains hotly debated28,29. Climate change generally leads to drier conditions and more ‘flammable days’ in dry woodlands10, and this is assumed to affect regional and global fire dynamics28,29. Longer and more frequent droughts are suggested to be driving increasing fire dynamics in Australia30 and South America31. However, many tropical dry woodlands, including these regions, are also hotspots of agriculture-driven deforestation19,32, potentially explaining more extensive and more frequent fires33,34. Moreover, there are potentially large but weakly understood interactions between climate and land use that affect fire patterns. For example, land-use actors may reduce the use of fire during drought periods to prevent runaway fires34 or, conversely, increase the use of fire for clearing land, controlling shrub encroachment or clearing crop residues to make use of favourable burning conditions during drought33,35. Given the often-detrimental outcomes of fires in dry woodlands, from environmental degradation36 to human health impacts37, disentangling the relationship between drought, land use and fire is important.

This, however, is non-trivial for several reasons. First, separating the roles of land use and climate in changing fire patterns in tropical dry woodlands requires multi-decadal time series of these variables, but such data are unavailable for most regions, particularly before 2000. Longer timescales are important for robustly capturing trends and baseline conditions prior to widespread land conversions or climate change38. Second, understanding the impact of land-use change on fire dynamics should move beyond only assessing initial deforestation39,40 towards analysing post-deforestation land use, which can strongly shape fire regimes41. Consolidated high-resolution satellite image archives from the Landsat and Sentinel sensor family42, cloud-processing capabilities43 and improved algorithms for working with very large volumes of satellite images now provide the possibility to address these knowledge gaps by reconstructing fire histories and land-cover changes over nearly four decades with unprecedented spatial and temporal detail44,45. This provides large but so far untapped opportunities to understand what drives fire dynamics in tropical dry woodlands.

Here we capitalize on these opportunities to analyse the relationship between fire, drought and land-use change for the entire 1.1-million-km2 South American Chaco, extending into Argentina, Bolivia and Paraguay (Fig. 1, small inset). The Chaco naturally consists of a mixture of dry forests, savannahs and grasslands; however, it has recently turned into a global hotspot of deforestation and land-use change46, with >35% of woodlands converted to agriculture since the 1980s (that is, ranching or cropping, mainly maize and soybean47). Drought, deforestation and land management in the Chaco have all been linked to fires48,49, and the region has recently been identified as a hotspot of increasing flammability potential due to climate change27. In this context, our goal was to reconstruct the Chaco fire history for the period 1985 to 2023, covering almost the entire history of industrial agricultural expansion in the region, and relate fires to indicators of deforestation, post-deforestation land use and drought. Specifically, we asked the following questions: (1) How have fire dynamics changed in the Chaco since 1985? (2) What is the relative importance of deforestation, post-deforestation land use and drought in explaining fire dynamics?

Fig. 1: Relationship between deforestation and fire in the Chaco.
Fig. 1: Relationship between deforestation and fire in the Chaco.The alternative text for this image may have been generated using AI.
Full size image

a, Proportion of area in 1 × 1 km2 cells that experienced woodland loss versus burning. The individual colours represent combinations of percentiles (for example, 20% and 40%) of each variable. ARG, Argentina; BOL, Bolivia; PRY, Paraguay. b, Percentage of fires related to clearing from 1986 to 2023 (that is, fire in natural vegetation with agriculture subsequently established). c, Percentage of fires related to management from 1986 to 2023 (that is, fires on croplands or pastures). d, Percentage of other fires from 1986 to 2023 (that is, fires in natural vegetation without subsequent agricultural expansion). Basemaps from GADM (https://gadm.org/) with data from the NASA Shuttle Radar Topography Mission, US Geological Survey, US Department of the Interior (https://doi.org/10.5066/F7PR7TFT).

Results

Reconstructing fire histories on the basis of ~175,000 Landsat images across the Chaco revealed distinct periods of higher and lower fire activity (Fig. 2). We found a clear peak in the mid-2000s: the annual burned area for the entire Chaco increased from the mid-1980s to around 2004, when it was largest (nearly 60,000 km2), with a subsequent decline to an overall low point in 2014 (~8,600 km2). This pattern broadly coincides with the overall trend in deforestation, with deforestation being particularly high during years of high fire activity (Fig. 2a), and this relationship was stronger when we looked at the entire Gran Chaco and at Argentina individually, compared with Bolivia and Paraguay, where our analyses point to higher statistical uncertainty (Fig. 2a). Burning rates generally followed this trend, and all three countries had particularly high rates in 2021/22. Our analyses also showed marked differences between countries, with Paraguay and Argentina following a similar overall trend (that is, burned area peaks in the 2000s). In contrast, Bolivia had low fire activity in the Chaco over the entire period we analysed, albeit with a marked recent increase, when deforestation in the Bolivian Chaco also increased (Fig. 2).

Fig. 2: Links between fire and deforestation.
Fig. 2: Links between fire and deforestation.The alternative text for this image may have been generated using AI.
Full size image

a, Relationship between annual burned area and deforestation in the same year. b, Trends of total burned area, burned area rates and deforestation (five-year averaged time series) for the period 1986–2023. c, Share of each fire type in total annual burned area. d, Anomalies of burned area and precipitation for each year in 1986–2023. Note that a–d show anomalies relative to their respective regional baselines. The plots representing the entire study region aggregate data across all subregions. As such, pronounced anomalies in a specific country may not be apparent in the overall regional plot. In a and d, we applied linear regression (ordinary least squares) to quantify the association; all P values are two‑sided, and no adjustments for multiple comparisons were applied. The shaded areas represent the confidence bands from the linear regression.

Chaco countries differed majorly when we overlaid burned areas with a detailed reconstruction of land-cover change for the same period47. This land-cover time series captured deforestation as well as post-deforestation land-use transitions (for example, changes from ranching to cropping) and allowed us to assess the proximate causes of fires (Figs. 1 and 2). According to this analysis, most fires (70% of all fires) in the Chaco were related to land clearing for expanding agriculture, particularly in Bolivia (91%) and Paraguay (83%). The second-most prevalent proximate fire cause was management fires (that is, fires on cropland or pastures, 28% of all fires in the Chaco), especially in Argentina (on average 34%), and to a lesser extent in Bolivia (7%) and Paraguay (16%; Fig. 2c). Management fires became more prevalent after 2015, accounting for 42%, 5% and 23% of all fires in Argentina, Bolivia and Paraguay during 2015–2023, respectively. Other fires, including other anthropogenic fires and natural fires in the remaining woodlands, occurred in all three countries (2%, 0.5% and 2% of all fires in Argentina, Paraguay and Bolivia, respectively) but have declined in importance in Argentina since the 2000s, when deforestation in the Chaco peaked. This trend was less pronounced in Paraguay, where fires had already decreased prior to the peak of deforestation (Fig. 2d). Other fires were important in Bolivia throughout our observation period until recently, but fires were overall less prevalent in Bolivia.

Combining fire histories with land-cover histories allowed us to analyse whether agricultural expansion (that is, the conversion of natural vegetation into pasture or cropland) or agricultural intensification (that is, the conversion of pasture to cropland) led to changes in fire activity. This showed that agricultural expansion strongly increased fire occurrences in the landscape. Where woodlands were converted to cropland, fire occurrence spiked strongly during the conversion period (from a median probability of ~25% to 87%), after which it declined to pre-conversion levels. A similar pattern occurred where woodlands were converted to pastures (the median fire occurrence increased from ~25% to ~70%; Fig. 3), highlighting the role of fire in the agricultural expansion process. Fire occurrence on pastures remained high and declined only slightly in the post-deforestation period, whereas when pastures were converted to cropland, fire occurrences dropped sharply, below levels characterizing unconverted woodlands (<20%; Fig. 3). This highlights the strong connection between management fires and ranching (but not cropping) in the Chaco.

Fig. 3: Fire occurrences before and after the three types of land-use change.
Fig. 3: Fire occurrences before and after the three types of land-use change.The alternative text for this image may have been generated using AI.
Full size image

Woodland–cropland conversions are presented only for Argentina and the entire Chaco, as not enough observations were available for the other two countries. The vertical purple lines represent the normalized year of the respective land-cover conversion, and the areas to the left and right refer to the immediate time window before (left) and after (right) conversion. Further information on the estimation of fire occurrence is provided in Methods.

To explore the relationship between fires and drought, we used a zero-inflated generalized linear mixed model to relate a fire indicator (total burned area) to a drought indicator (drought intensity). This showed that increasing drought intensity (that is, a stronger annual precipitation deficit) was strongly related to increasing numbers of fires (Fig. 4). This increase was particularly pronounced for clearing and management fires, where the burned area increased strongly with higher precipitation deficits. The trend was strongest for clearing, where the proportion of burned area during drought years (with a normalized precipitation anomaly of 1) was nearly 5% higher than during non-drought years (with a normalized precipitation anomaly of 0). Surprisingly, the burned area associated with other (non-land-use) fires declined with an increasing precipitation deficit (Fig. 4).

Fig. 4: Number of fires and burned area type versus precipitation anomaly.
Fig. 4: Number of fires and burned area type versus precipitation anomaly.The alternative text for this image may have been generated using AI.
Full size image

Left: number of fires versus precipitation anomaly. Each dot in the graph represents the summary for an individual year across the entire study area. The number of fires was extracted from our burned area maps as individual fire patches, independent of the causes of fires. Right: burned area type versus precipitation anomaly. The graph represents the regression outcomes of our zero-inflated generalized linear mixed model, which we ran at a spatial resolution of 1 × 1 km2. The shaded areas represent the confidence bands on the regression lines. The axes and units are therefore interpretable as densities (dimensionless) and as areas (km2).

Discussion

Many tropical regions in South America22 and elsewhere11 have recently experienced record fire years, and these increases in burning are frequently attributed to climate change, specifically to the occurrence of longer and more severe droughts27. Simultaneously, there are signs that many of these fires are anthropogenic and associated with land management33,34. Given the detrimental environmental, social and human health outcomes of fires, disentangling the roles of climate and land use in driving fire patterns is critical for designing adequate policy responses to prevent fires. This is particularly important for the South American dry woodland regions, such as the Caatinga, Cerrado, Chiquitano and Chaco, many of which experience an increasing frequency and severity of droughts27, rampant agricultural expansion and intensification50, and increased burning28. Focusing on the entire 1.1-million-km2 Chaco region, a global deforestation hotspot extending into Argentina, Bolivia and Paraguay, we show that the majority of fires are attributable to agriculture and that such fires become particularly frequent during drought. This challenges the notion that drought is a principal driver of more widespread and severe fires in tropical South America. Instead, our findings suggest that land use appears to be the leading cause of increasing fires in the Chaco and possibly other tropical dry woodlands.

Our results show that fire is indeed an important process, with fires widespread throughout our observation period and a total area of more than 70 million hectares affected by fire since 1986, corresponding to nearly two thirds of the Chaco experiencing burning at some point since the 1980s. Fires peaked during the mid-2010s and declined thereafter before becoming more widespread in recent years. Three independent lines of argument suggest that intensifying fire regimes in the Chaco are driven by land use. First, overall trends in burning over time coincided with overall trends in deforestation, with the peak of fire activity in the Chaco corresponding well with the peak of deforestation in the early and mid-2000s47 (Fig. 2a,b). Examining post-deforestation land uses provides further evidence of the importance of agriculture in explaining fire activity. Among land-use-related fires, management fires dominated, most of which were associated with pasture management (Fig. 3). This coincides well with what is known about the region. Fires are heavily used for controlling woody encroachment on pastures, a major factor decreasing the productivity of ranching in the region51. On the other hand, fires are known to be rarely used for removing crop residues in the Chaco, as much of the cropping there is based on zero-tillage with alternating low-biomass (soybean) and high-biomass (for example, maize) crops to sustain soil fertility via organic matter inputs. Crop residues (that is, the barbecho) are therefore intentionally not burned. Additionally, fires on croplands are forbidden (that is, under Law 26.815 in Argentina)21 or subject to strict controls (that is, Laws 4014/2010 and 294/1993 in Paraguay), explaining why they are virtually absent in the Chaco. We highlight that a considerable share of ‘other’ fires we mapped (that is, those not occurring in agricultural areas or where deforestation leads to agriculture) might still be linked to land use—for instance, when fires used for pasture management spiral out of control and extend into adjacent forests52.

Second, fire extent and frequency were particularly high during intensive drought years (Fig. 2d), confirming previous findings for the Argentinian Chaco48,51,53,54 and Bolivia55. While at first glance this suggests a link between extreme meteorological conditions and fire activity, by examining the interactions between land use, drought and fire, we empirically show that a large extent of this increase is linked to fires connected to land clearing or land management. We suggest that this is because drier conditions motivate actors to use fire to clear land cheaply and to clear more land than under wetter conditions, a phenomenon suggested for Kalimantan’s peatlands56 and Amazonia33. In the Chaco, which has been subject to more frequent and prolonged droughts in recent decades27, this may also be the case. Likewise, in drier, low-productivity ecosystems, such as the Chaco, natural fires may become fewer with increasing fire weather severity due to diminishing fuel loads57, corresponding well to our models that show fewer ‘other’ fires during longer drought conditions. Our results thus contradict recent global assessments of links between fire and climate change, which for dry woodlands mainly attribute evapotranspiration and vapour moisture deficit as drivers of fire patterns28,58, neglecting the roles of land use and people59.

Our third line of argument uses the natural experiment that the Chaco, extending into three countries, presents: we found clear country-specific fire dynamics that should not occur under a general climate–fire regime without the land-use linkage. For example, in Argentina, the burned area and deforestation changed overall in parallel, with both peaking around 2005 and a decreasing trend thereafter. However, in Paraguay most deforestation happened during a period of relatively low fire activity between 2010 and 2015. Management fires on pastures were most common in Argentina, where fires are frequently used for shrub removal and to promote palatable grass regrowth in both pastures and silvopastures60. Conversely, in Paraguay, where ~90% of the agricultural expansion was for ranching61, the most important strategies to maintain and improve pasture productivity involve the intermittent planting of nitrogen-fixing legumes and soil-covering techniques62 as well as mechanical shrub removal, explaining the lower share of pasture management fires we found there (Fig. 2). While some uncertainties remain in our satellite-based time series of burning, deforestation and land-use change, this overall relatively low uncertainty (Supplementary Text 1) should not be systematically biased in space, in time or among countries. Together, this suggests that fire dynamics in the Chaco are driven by land use, with drought as a mediating factor.

Fire is a natural component of tropical dry woodlands; however, fire regimes are changing in the Anthropocene in many of these regions. Climate change is expected to lead to hotter and drier conditions across the tropics28, which is expected to intensify fire regimes. Here we found that the link between fire and climate change might be mediated by land use in major ways. Fires connected to agricultural expansion and pasture management were the dominant types of fires in the Chaco, and these fires became both more frequent and more widespread during drought years. This suggests that considering and disentangling the drought–fire–agriculture nexus is necessary for understanding the social ecology of contemporary, increasingly human-dominated tropical dry woodlands32, where most of the research on fire has either been broad-scale or focused on natural ecosystems (for example, inside protected areas), both largely disregarding land use and people59. Currently, 40% of tropical forest loss globally is related to fires17, and approximately a quarter of all tropical fires are found in croplands23. These fires are of major concern in terms of carbon emissions7, biodiversity losses12 and human health impacts63. Our finding that fires in tropical dry woodlands are largely caused by land use and not climate extremes provides important opportunities to better manage fires and avoid the social-ecological costs they incur. Leveraging these opportunities probably requires a combination of (a) stricter deforestation policies and enforcement; (b) improved fire management, particularly during drought years, and related education measures; and (c) better fire monitoring and adequately funded rapid-response mechanisms to prevent fires from developing and spreading—in the Chaco as well as in other tropical dry woodlands.

Methods

Reconstructing fire histories

We reconstructed fire history in the Chaco by mapping burned area annually between 1986 and 2023. Specifically, we used all available Landsat satellite collection 2 tier 1 surface reflectance data to calculate 80 spectral temporal metrics from the multispectral bands and a set of spectral indices (NBR, NBR2, NDMI, NDWI, EVI and the three tasselled cap components brightness, greenness and wetness) for each phenological year (that is, March to March) of the study period (1986–2023). We collected training data for fires that occurred through on-screen digitization and broadly identified two types of fires: (1) fires in woodlands and (2) agricultural fires, which we identified using normalized burn ratio composites and MODIS active fire data as a reference. In total, we collected 822 training polygons from 30 years (see Supplementary Fig. 1 for a temporal distribution of these files). We then sampled 10,000 random points each inside the polygons and across woodland and agricultural areas in the Chaco. This dataset provided the basis for our analysis. We then parameterized a time-calibrated random forest model to calculate the burned area probabilities for each pixel and each year from 1985 to 2023. We converted the continuous burned area probabilities into binary burned area maps by identifying pixels of very high burn probability (that is, >95%) and then using a random walker segmentation64 to identify burn scars in the landscape44. This provided, for each year in 1986–2023, a binary map of burned versus unburned areas. The mapping of the burned area probabilities was done in Google Earth Engine using the Python interface (Python version 3.10) (ref. 65), and the random walker segmentation was done in Python using the scikit-image library (version 0.25.2).

We rigorously validated our maps using an independent validation sample, following best practices66 (for more details on the validation, please refer to Supplementary Text 1). This showed that our annual burned area classification had a mean overall accuracy of >94% across all years (that is, the mean of all yearly overall accuracies). We then calculated the burned area statistics for each year for the entire Chaco, as well as for the Chaco sections in the three countries (Argentina, Bolivia and Paraguay). Further information on the burned area mapping is provided in Supplementary Text 1. Using dense time series of burned area and land-cover change derived from the entire Landsat satellite record allowed us to analyse the links between burning and land use in unprecedented spatial (that is, 30 m) and temporal (that is, yearly) detail over a nearly 40-year period. Our fire history was highly reliable (overall accuracy >94%, F1-score 43% for the burned area class and 93% for the non-burned area class). Similarly, our land-cover-change time series was highly reliable (overall accuracy >90%), especially for the woodland and agricultural classes on which our work was based (user’s and producer’s accuracies on average >88%).

Relating fire dynamics to land-cover change

To understand the role of land use in shaping fire dynamics, we combined our burned area time series with a consistent land cover time series for the Chaco from our prior work47. By overlapping these time series, we identified and mapped three categories of fires: (1) clearing fires, representing fires occurring in natural vegetation (that is, woodlands, savannahs or natural grasslands) where agriculture (that is, croplands or pastures) was established after the clearing (at the latest three years after the fire); (2) management fires, representing fires on existing croplands or pastures; and (3) other fires, representing fires in natural vegetation areas that were not converted to agriculture in the next three years. These fires include both natural fires and other anthropogenic fires. We summarized these fire categories for the entire Chaco, for the dry and wet Chaco separately and for each country for the period 1986–2023.

We also assessed whether agricultural expansion (that is, the conversion of natural vegetation to pasture or cropland) or agricultural intensification (that is, the conversion of pasture to cropland) led to an increase or decrease in fire activity. For this, we aggregated our 30-m land-cover-change maps to 3-km cells (that is, 100 × 100 pixels in our native resolution) by calculating the area cleared for cropland and pasture for each year, as well as the area of pasture that was converted to cropland. Next, we examined the fire history before and after conversion for all converted areas by calculating the proportion of the converted area that had burned in a given year and defined this as fire occurrence. For example, for a given 3 × 3 km2 cell, we assessed all areas converted from woodlands to pasture in 2000 (year 0 in this example) and assessed the share of these areas that burned in 2000, as well as for 10 years before and after 2000 (so 1990–1999 and 2001–2010 in our example). This allowed us to track changes in fire occurrence on the same piece of land before and after land-use change occurred41. We then merged the results of all 3-km cells by calculating the mean fire occurrence across all cells for each land-cover conversion individually. For woodland-to-pasture and pasture-to-cropland conversions, we did this for the entire Chaco and the three individual countries. For woodland-to-cropland conversions, we calculated fire occurrences for the entire Chaco and Argentina, but not for Bolivia and Paraguay as there were not enough fire observations there. This approach allowed us to compare changes in fire occurrences across space and time, as our calculations were independent of a specific time window.

Relating burned area to drought

To better understand the effect of droughts on fire dynamics, we gathered CHIRPS precipitation data67 and calculated annual precipitation anomalies relative to the long-term (that is, 1981–2010) mean annual precipitation for 1-km grid cells. Next, we calculated the area of our three fire types (that is, clearing, management and other fires) for each year and cell. To estimate the effect of precipitation anomaly on the burned area, while accounting for fire type, we then combined these data in a zero-inflated generalized linear mixed model68, in the form of

$${{\rm{BA}}}_{{\rm{id}},{\rm{yr}}}={{\rm{Anom}}}_{{\rm{yr}}}\times {\rm{Type}}+(1|{\rm{id}})+(1|{\rm{yr}})$$

where BAid,yr represents the burned area in cell id and year yr, the term Anom × Type represents the interaction between our drought variable (that is, precipitation anomaly) and fire type, and (1|id) and (1|yr) are random effects for each cell and year, respectively. We ran the glmmTMB function of the glmmTMB package in R (version 1.1.14)68 assuming BA follows a Tweedie distribution (that is, a distribution suitable for continuous data that include many zeros).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.