SCIENCE ADVANCES | RESEARCH ARTICLE APPLIED ECOLOGY Copyright © 2021 The Authors, some Fire-induced loss of the world’s most biodiverse forests rights reserved; exclusive licensee in Latin America American Association for the Advancement of Science. No claim to Dolors Armenteras1*, Liliana M. Dávalos2,3, Joan S. Barreto1, Alejandro Miranda4,5, original U.S. Government Angela Hernández-Moreno6, Carlos Zamorano-Elgueta4,7, Tania M. González-Delgado1, Works. Distributed María C. Meza-Elizalde1, Javier Retana8 under a Creative Commons Attribution Fire plays a dominant role in deforestation, particularly in the tropics, but the relative extent of transformations and NonCommercial influence of fire frequency on eventual forest loss remain unclear. Here, we analyze the frequency of fire and its in- License 4.0 (CC BY-NC). fluence on postfire forest trajectories between 2001 and 2018. We account for ~1.1% of Latin American forests burnt in 2002–2003 (8,465,850 ha). Although 40.1% of forests (3,393,250 ha) burned only once, by 2018, ~48% of the ever- green forests converted to other, primarily grass-dominated uses. While greater fire frequency yielded more trans- formation, our results reveal the staggering impact of even a single fire. Increasing fire frequency imposes greater risks of irreversible forest loss, transforming forests into ecosystems increasingly vulnerable to degradation. Reversing this trend is indispensable to both mitigate and adapt to climate change globally. As climate change transforms fire regimes across the region, key actions are needed to conserve Latin American forests. INTRODUCTION Fire can be essential to many ecosystems globally, as it consistently Forests cover nearly one-third (4.06 billion ha) of land globally (1) influences ecosystems along multiple dimensions ranging from physico- and play a central role in the earth system by providing global eco- chemical properties to biotic patterns and processes (7). Since fire system services such as climate regulation and water cycling (2) and can also cause critical shifts, it has been identified as a prevalent as reservoirs of land biodiversity. More than half (56%) of world forest disturbance in the tropics causing forest degradation (1, 8). forests are tropical (45%) and subtropical (11%) (1), and these are Increasingly, many drivers of land use change interact and operate also the forests experiencing most deforestation. Globally, annual synergistically (9) and quickly (3). Global warming caused by cli- net forest loss dropped by 20% (10 million ha) in 2015–2020 com- mate change increases forest vulnerability to fire through different pared to 2010–2015 (12 million ha). Despite this decline in global pathways. First, global warming alters fire regimes by increasing the forest loss, forest area is still decreasing (3): Africa had the largest duration, frequency, and intensity of the dry season. These climatic forest loss in 2010–2020 (3.9 million ha), followed by South America alterations are strongly related to the El Niño-Southern Oscillation (2.6 million ha) (1). and ocean-surface temperature anomalies (10), directly increasing Tropical forest loss is associated with human activities that con- the occurrence of wildfires and subsequent forest die-offs (11). Sec- tinuously threaten these ecosystems, from ranching to infrastructure ond, extreme events such as droughts cause larger, more frequent, expansion, causing biodiversity loss and decreased soil and water intense, and severe fires resulting in higher carbon emissions per protection. Although the globally burnt area declined between 1998 burned unit area (11). Droughts increase forest inflammability (12) and 2015, increases in fire frequency and intensity make this driver through fuel accumulation on the forest floor or increases in the air of forest conversion and degradation crucial to understanding de- temperature. The interaction of climate change with natural topog- forestation (4). While this reduction in burned area (BA) was mainly raphy and wind conditions, as well as other anthropogenic activities, associated with tropical savannas and grasslands (5), the global BA can then result in severe and extended fires, further affecting large has shifted to regions with more tree cover: closed canopy forests. forest areas and reducing forest carbon storage (11–13). Fire-induced forest This shift has been traced to the use of fire for deforestation and degradation in tropical regions can represent close to 69% of their agricultural management in the tropics (5). As a result, from 2001 to total carbon lost (14). Together, deforestation and degradation of forests 2015, wildfires contributed to 23% (± 4%) of forest loss (4) and a in the American tropics have led to losses of approximately 516 ± carbon emission of 454 ± 496 Tg CO2 year−1 worldwide (6). 69.5 Tg C year−1 (14), indicating that most losses are associated with fires. Latin American forests are of planetary importance, with South America alone holding 21% of the world’s remaining forests (15). While both forest cover and naturally regenerating forest area have 1 Laboratorio de Ecología del Paisaje y Modelación de Ecosistemas ECOLMOD, De- declined in South America since the 1990s (1) resulting in 7% of re- partamento de Biología, Facultad de Ciencias, Universidad Nacional de Colombia, Sede Bogotá, Bogotá, Colombia. 2Department of Ecology and Evolution, Stony Brook gional forests lost in under 15 years (1990–2005) (16), these dynamics University, 630 Life Sciences Building, Stony Brook, NY 11794, USA. 3Consortium for prevail across rainforests worldwide (17). Over the past decade, and Inter-Disciplinary Environmental Research, School of Marine and Atmospheric Sci- thanks to Brazil’s ability to reduce Amazon Basin deforestation rates ences, Stony Brook University, 129 Dana Hall, Stony Brook, NY 11794, USA. 4Center for Climate and Resilience Research (CR2), Santiago, Chile. 5Laboratorio de Ecología (17), forest loss in the Amazon declined from peaks in the 2000s. del Paisaje y Conservación, Departamento de Ciencias Forestales, Universidad de However, political and economic moves to undermine forest con- La Frontera, Temco, Chile. 6Centro de Investigación en Ecosistemas de la Patagonia servation in Brazil have since brought back massive deforestation in (CIEP), Camino Baguales s/n Km 4, Coyhaique, Chile. 7Departamento de Ciencias both “legal Amazonia” and elsewhere in the basin (18). Naturales y Tecnología, Universidad de Aysén, Coyhaique, Chile. 8CREAF- Universitat Autonoma Barcelona, 08193, Cerdanyola del Valles, Barcelona, Spain. In the Amazon, forest degradation from the extraction of forest *Corresponding author. Email: darmenterasp@unal.edu.co resources (19) is leading to major forest and biodiversity loss. A number Armenteras et al., Sci. Adv. 2021; 7 : eabd3357 13 August 2021 1 of 8 SCIENCE ADVANCES | RESEARCH ARTICLE of factors drive land use change, but conversion of forests to pastures of forests to increased frequency of fires, and those available are largely for cattle ranching and agriculture is prevalent in Latin America (9). based on local observations (31). Systematic analyses of fire-associated Bolivia, Peru (20), and Colombia (21) follow the Brazilian pattern, forest loss, examining forest trajectories under changing fire frequency, wherein most forest loss is caused by agricultural expansion (20). can help disentangle the mechanisms determining forest degradation Fire is routinely used to convert forests to agricultural uses and has and examine forest capacity for resilience. Here, we address three been identified as a direct driver of deforestation in the region (22). key questions. First, what proportion of 2001 Latin America forests, Fires used for clearing vegetation or burning agricultural fields or including but not limited to the Amazon region, burned in 2002/2003? pastures often result in unintended wildfires that spread into adja- Second, what are the trajectories of postfire change in those burnt cent forests, degrading them (23). Once the agricultural frontier forests? Third, how does fire frequency influence those trajectories? consolidates, fire activity tends to decline (24), but the cycle of col- We draw on advances in remote sensing to discover forest areas af- onization and consolidation spills into nearby forests where pastures fected by fire in 2002/2003 and follow their land cover trajectory up and crops will then be established, with subsequent deforestation to 2018, providing methods to account for the role of fire frequency on and expanding fire-­related degradation. Wildfires, in turn, interact forest conversion in Latin America and beyond. Specifically, we analyzed with other drivers of tropical forest loss such as fragmentation (25), data for forests in 2001 using previously published moderate resolu- selective logging, or even policy changes (26) that, through complex tion imaging spectroradiometer (MODIS) land cover and burnt area feedbacks, increase the vulnerability of remaining forests to further datasets (32), assessing how fire frequency influenced forest conver- degradation. Thus, exposure and altered microclimates in forest sion and modeling land cover trajectories in the region. edges result in increases in forest fire occurrence, magnifying sub- sequent impacts (27). Already altered tropical forests are less resistant to climate change RESULTS and human activity (13). Once disturbed, forests are even more vul- Area burned (2002/2003) nerable to other disturbances and subsequent fires (6). Increased Latin America had 7,781,832 km2 of forests in 2001 (Fig. 1A), mostly susceptibility of tropical forests to fires has resulted in a spike in evergreen forests (EF; 7,194,695 km2), deciduous broadleaf (DB; fire-related loss, but remaining forests have also become more vul- 446,212 km2), and mixed forests (MF; 140,925 km2). The total land nerable to fire under drier conditions (3). Forests in South America area burned in 2002/2003 was 84,451,200 ha, of which more than are predicted to have low resilience, particularly in areas with a high half was savannas [45.56% savannas and 8% woody savannas (WS)], risk of future droughts (28). Forest fire studies often focus on mea- followed by grasslands (26.64%) and 10.65% forests. In response to suring forest area burnt (29) or its effect on carbon cycling and atmo- our first question, a total of 8,465,800 ha of 2001 forests were affected spheric emissions (30). Yet, few studies have analyzed the vulnerability by fire in 2002/2003, so 1.09% of Latin American forests were burned Fig. 1. The magnitude of burning occurrences per land cover in Latin America. Distribution of (A) land cover types in Latin America and (B) selected forest pixels with fire frequency between 2002/2018 (one, two, and three or more occurrences). Armenteras et al., Sci. Adv. 2021; 7 : eabd3357 13 August 2021 2 of 8 SCIENCE ADVANCES | RESEARCH ARTICLE that period. A 74.4% of the affected forest were EF (6,298,100 ha), forest pixels burnt twice corresponded to 77.8% (i.e., 1,466,625 ha) followed by a 23.4% of DB forests (1,977,100 ha) and the rest corresponded were EF, and 20.3% (382,275 ha) and the rest 2% (36,775 ha) were to MF (190,500 ha). The forests burned represented ~1.09% of all MF. Those pixels burnt thrice, 66.4% (i.e., 2,052,375 ha) were EF, and 32.1% Latin American forests in 2001. (1,023,775 ha) and the rest 3.4% (109,150 ha) were MF (Table 1). Globally, Brazil has the largest extension of forest and had the most forest area affected in 2002/03 with 4,695,775 ha of forest affected Proportion of land cover types and trajectories by fires (1.2% of total forest present in 2001). Similarly, in Bolivia, of postfire change 1,431,325 ha of 2001 forests were burnt (2.6% of total forest) while Overall, and in response to our second question, the trend of forest in Mexico with 983,325 ha burnt (3.4% of total forest in 2001). Paraguay loss over time is clear, forests are mostly replaced by savannas and (8.4%), Guatemala (7.8%), and Honduras (6.1%) had less total area grasslands. Both regional and cross-country trajectories of postfire affected but a greater proportion of their forest burned (table S1). changes indicate that in all cases, even after being burnt only once and following the initial loss in forest cover of 2001, there is a general Frequency of fire (2002/2018) declining trend of proportion of forest cover over later years. How- Approximately 40% of forests pixels that burned in 2002/2003 ever, there is some variation in temporal patterns (fig. S1). For some (3,393,250 ha; Fig. 1B) then did not undergo any other fire event in countries, forest loss trajectories after fire are faster in the years im- the following 15 years (i.e., fire frequency of one up to 2018). Around mediately after the 2002/2003 fire. While Argentina, Brazil, Chile, 1,885,675 ha (22.4% of the forests affected in 2002/2003) experienced Guyana, and Panama had rapid losses, Mexico, Guatemala, Honduras, a second fire (Fig. 1B), and the remaining 3,186,925 ha (37.6%) un- and Venezuela had slower declines in the proportion of forest cover derwent three or more fires up to 2018 (Fig. 1B). over the years (fig. S1). In Chile, forest recovered after 2004, for Of the forest pixels burned only once, 81.9% (i.e., 2,777,675 ha) were Surinam regrowth followed 2009, and Colombia had a small increase EF, and 16.8% (570,925 ha) and the rest 1.3% (44,650 ha) were MF. Those in forest cover after 2017. In response to our third research question, the overall proportion of forest loss varies considerably with changing fire frequency (Fig. 2). For forest pixels undergoing only one fire, the proportion of forests Table 1. Area of forest burnt in 2002/2003 per forest cover type and fire frequency. is reduced to 74% by 2004, the year after the initial fire. This reduc- tion in the proportion of forest is more evident for EF, which re- Burned once Burned twice Burned thrice duced their proportion to 58.8% by 2004 and further to 43.5% by Total forest (ha) 3,393,250 1,885,675 3,185,300 2018. In forests that burned twice during the period of analysis, the EF 81.9% 77.8% 64.4% proportion of forests reduces initially to 66.4% by 2004, and again, (ha) 2,777,675 1,466,625 2,052,375 EF fall from to 47.8% by 2004 and further to 26.7% by 2018. These DB forests 16.8% 20.3% 32.1% trajectories are more marked in those forests burned thrice or more: A mere 14% of those EF remain by 2018. (ha) 570,925 382,275 1,023,775 MF 1.3% 2% 3.4% Land cover type changes per fire frequency (ha) 44,650 36,775 109,150 Greater fire frequency undermines forest persistence from 2001 to 2018 (Fig. 3). Although most forest loss concentrates in the 2001–2004 100 75 Once 50 25 0 Cover 100 Percent cover Evergreen forests Deciduous broadleaf forests 75 Mixed forests Twice 50 Woody savannas Savannas 25 Grasslands 0 Croplands 100 Other Thrice or more 75 50 25 0 01 02 03 04 05 06 07 08 09 10 16 17 18 11 12 13 14 15 20 20 20 20 20 20 20 20 20 20 20 20 20 20 20 20 20 20 Fig. 2. Proportion (%) of land cover types over time per fire frequency. One fire (top), two fires (middle), and three or more fires (bottom) in the period 2002–2018. Armenteras et al., Sci. Adv. 2021; 7 : eabd3357 13 August 2021 3 of 8 SCIENCE ADVANCES | RESEARCH ARTICLE Gain Loss Stable cover Stable type Once Twice Thrice or more 100 Evergreen forests 75 Deciduous broadleaf forests Percent land cover Mixed forests Woody savannas Savannas 50 Grasslands Croplands Other 25 0 2001 2004 2018 2001 2004 2018 2001 2004 2018 Fig. 3. Land cover type trajectories between 2001 and 2018 for forests with frequency of one, two, or three or more fire events. The width of the land cover type bar in each case represents the proportion of the total forest area burnt. The color of the trajectories indicates whether the land cover type changes to another land cover type (gray), maintains stable (green), the forest type changes but forest cover remained (pink), or the change indicates some recovery (blue). Loss Stable forest cover Stable forest type Arg Bel Bol Bra Chi Col Gua Guy 100 75 Evergreen forests 50 Deciduous broadleaf forests Percent land cover 25 Mixed forests Woody savannas 0 Savannas Hon Mex Nic Pan Par Per Sur Ven Grasslands 100 Croplands Other 75 50 25 0 01 18 01 18 01 18 01 18 01 18 01 18 01 18 01 18 20 20 20 20 20 20 20 20 20 20 20 20 20 20 20 20 Fig. 4. Land cover type trajectories in each country between 2001 and 2018 for forests with frequency of one, two, or three or more fire events. The width of the land cover type bar in each case represents the proportion of the total forest area burnt, with all countries but Brazil magnified by a factor of 2. The color of the trajectories indicates whether the land cover type changes to another land cover type (gray), maintains stable (green), or the forest type changes but forest cover remains (pink). period in all cases, subsequent years also show continuous forest with most 2001–2004 forest conversion to savannas (11.93%), while loss over time. Detailed analyses of pixel fate through time show a for the 2004–2018, remaining EF were lost to WS (16.5%). Patterns large portion of burnt 2001 forest converts to other nonforest land of forest loss and land cover type conversion for forest with higher cover types. This pattern especially holds for EF, the forest type com- fire frequencies reveal even more marked losses, with lower propor- prising the largest area burnt. tions of forest remaining stable at greater fire frequency. For EF burned once (2,777,675 ha), 27.3% (759,261 ha) was lost Within region and cross-country variation between the 2001 and 2004, just after fire, and a further 21% (581,830 ha) A general pattern of loss of forest cover after fire is found in all coun- of the 2001 EF area was lost between 2004 and 2018. The trajectory tries (Fig. 4). Belize had the highest forest persistence. In contrast, of loss reflects conversion mostly to WS, savannas, and grasslands, Paraguay, Panama, and Brazil had higher proportions of forest loss Armenteras et al., Sci. Adv. 2021; 7 : eabd3357 13 August 2021 4 of 8 SCIENCE ADVANCES | RESEARCH ARTICLE following fire. For EF, Chile and Belize had the highest forest persistence of both. Previous studies of postfire vegetation dynamics have found (83.9 and 82.5%, respectively). Panama (64.7%), Paraguay (61.5%), most South American forests can be severely affected by single fires and Brazil (56.6%) showed a relatively high proportion of EF con- and therefore recover only very slowly, requiring up to two decades version to other land cover types (Fig. 4 and table S2). In Panama, to grow back (35). While statistical analyses confirm persistence is EF converted mostly to savannas (48.6% WS, 11.6% S), as was the frequent enough after a single fire (Fig. 5), conversion to various case in Paraguay (33.8% S, 11.6% WS). Brazil EF converted to both nonforest cover types is at least as frequent, particularly for EF and savannas (23.4% S, 12.5% WS) and grasslands (15.9%). While EF in MF. This high variability in the frequency of transitions, together with Bolivia persisted at high proportion (70.7%), they also experienced among-country differences in the types of forests initially burnt (fig. S2), the highest conversion to croplands of all countries, at 5.5%, followed suggests fire is used in country-specific ways. Nevertheless, fire gen- by Brazil (3.5%) and Paraguay (3.1%). erally leads to cleared forests instead of to other forest cover types. For DB forests, Chile (78.9%) and Paraguay (68%) had higher We acknowledge that burnt forests that end up being deforested could forest conversion rate to other land cover types. Chile’s DB converted, result from several confounding effects that drive fire frequency and to a great extent, to croplands (26.8%), savannas (22.1% S, 11.3% WS), vary regionally such as biomass, ignition season, and length of dry and grasslands (11.3%). Argentina’s DB forests followed Chile’s con- season. Yet, our results highlight the markedly increased threat of version to croplands with 17.6% of the forests affected by fire in 2001. fires to Latin American forests: Even a single fire—we cannot ascertain Mexico and Bolivia’s DB forests, on the other hand, persisted better whether used as a human tool or wildfire—coincides with widespread after fire (88.7 and 80.3%). MF persisted better in Mexico (56%) and conversion to other uses. Despite forest persistence, areas that burned Chile (38.4%), but the Chile case also changed a high proportion of are also as likely to become deforested as not (Fig. 3) (36). these to grasslands (19.5%). Argentina’s MF converted mostly to Our analyses reveal how fire promotes the degradation of forest grasslands (68.2%) and savannas (15.9% WS, 6.8% S). into savannas and grasslands. Although the land cover types we an- Modeled influence of starting cover type and country alyzed cannot determine whether particular covers result from hu- To model the varying trajectories of the single-fire pixels and quan- man use, resulting savannas are primarily degraded forests with less titatively answer our second question, we adopted a hierarchical tree cover, instead of old-growth savannas, and grass-dominated Bayesian approach to estimate the frequency of changes from various vegetation in many cases reflects conversion to pastures for livestock forest cover types while accounting for country-specific variation in grazing and not natural grasslands (37). Forest-to-savanna land cover both the total pixels and the proportion of forest types burnt. All type transitions were indeed frequent, with transitions to crops at models converged, yielding posterior effective sampling sizes >200 much lower frequency, at least at this spatial scale (Fig. 5). Instead for all parameters. Deviation information criterion (DIC) compari- of smooth transitions to savannas and prairies, these were abrupt, sons (table S3) showed transition types differ (i.e., the best fit model directly following fires in EF and DB forests (Fig. 3). These sharp included interactions between start and end land covers). Country-­ transitions are both a response from fire-vulnerable species (12) specific effects were allowed to vary such that country identity ex- and the result of altered fire regimes (28) and conversion to pastures plained different amounts of variation depending on the starting forest type and initial forest cover types covary to different degrees. Sample-wide coefficients for this model (Fig. 5) reflect starkly vary- ing transition frequencies after accounting for country-specific vari- EF−>EF ation in initial forest cover type burned (fig. S2). Full model parameters EF−>DB EF−>MF are summarized in table S4. EF−>WS EF−>S EF−>G EF−>C DISCUSSION EF−>Other Blazing for weeks and emitting smoke that reached as far as São DB−>DB DB−>EF Paulo, the 2019 fires in the Brazilian Amazon highlighted the im- DB−>MF portance of anthropogenic fires in forest loss on a planetary scale. DB−>WS DB−>S Our results reveal less marked, but similarly consequential, fire events DB−>G have shaped forest conversion across Latin America in the Amazon DB−>C DB−>Other and beyond over the past two decades. These hitherto underexplored MF−>MF events both reinforce the links between fire and forest loss and re- MF−>EF veal the vast negative impacts that fire exerts over Latin American MF−>DB MF−>WS forests. While contributions to carbon emissions are often highlighted, MF−>S fire-sensitive EF were most affected, and species therein lack adap- MF−>G MF−>C tations to tolerate fires (12, 33). Thus, fires that obliterated half the affected MF−>Other forests over the past two decades have impacts far beyond biomass loss (34), threatening species, phylogenetic, and functional diversity. −10 −5 0 5 10 15 Shockingly, a single fire event can set the stage for a massive pro- portion of forest loss and its conversion into other uses. This asso- Sample-wide transition coefficients ciation may indicate that fires are agents of loss (i.e., humans use Fig. 5. Posterior distribution of sample-wide coefficients on transition fre- fire to deforest), a common cause drives more fires and deforesta- quencies for the single fire data. Posterior distributions whose 95% credible in- tion (e.g., climate change makes wildfires more common and peo- tervals do not overlap with 0 are shown in black, in gray those whose 50% credible ple opportunistically deforest the burnt area), or some combination intervals do not overlap with 0. Armenteras et al., Sci. Adv. 2021; 7 : eabd3357 13 August 2021 5 of 8 SCIENCE ADVANCES | RESEARCH ARTICLE (38). Burned tropical forest can become highly degraded systems from 2002 for all pixels that burned in the 2002–2003 period. Nei- and, in some cases, experience a transformation to human-derived ther 2002 nor 2003 experienced anomalous climate. savannas or other anthropogenic and novel ecosystems (37). Further, Here, we used the “Cover Type 1” band from the MCD12Q1 positive feedbacks between climate change and human activities then MODIS product that refers to a classification of land cover types change the structural and compositional trajectories of forest suc- using the IGBP (International Geosphere-Biosphere Program) meth- cession, driving further forest conversion and increasing fire prob- odology. This product has been widely used as a reference to obtain abilities at broader scales (39). After forests burn, the recovering land cover data (45); to study relationships between BA, forest loss, vegetation can become more vulnerable to climate variation and and cover changes (51); to analyze the relationship between climate surrounding pressures. For example, at the community level, forests change, fires, and changes in vegetation (52); or to analyze BA in in postfire recovery stage comprise fast-growing pioneer species, relation to cover types for the Global Fire Atlas (47). We reclassified typical in early successional stages, which can increase drought sus- the original category into 8 new land cover types by maintaining ceptibility, thus increasing fuel accumulation and vulnerability to some of the 17 original land cover types: DB forests (type 4, defined fire (40). as forest cover dominated by DB trees with canopy >2 m and tree Although fragmentation-fire feedbacks were not directly exam- cover >60%), MF [type 5, defined as forests dominated by neither ined here, fragmented forests are more exposed to anthropogenic deciduous nor evergreen (40 to 60% of each) tree type (canopy >2 m) fires (39). Edges created by fragmentation favor hostile climatic and tree cover >60%], WS [type 8, areas with tree cover 30 to 60% conditions that influence vegetation regeneration and persistence (canopy >2 m)] and savannas (S; type 9, defined as areas with tree (40). Since fragmented forests have higher edge-to-core ratio, they cover 10 to 30% (canopy >2 m)], grasslands [type 10, dominated by are more vulnerable to fire intrusion, and further synergies between herbaceous annuals (<2 m)], or croplands (type 12, pixels with at fragmentation and fire ensue (12, 25, 41). Together with shifts in least 60% of area is cultivated cropland). One key limitation of this community composition, these physical changes make subsequent classification is the inability to discern human use for grass-dominated fires burn hotter and longer than in intact forests. Our analyses un- vegetation, as natural grasslands and pastures (with anthropogenic covered repeated fires are a powerful determinant of forest conver- use) are merged for the grassland land cover. sion and there is greater forest loss with higher fire frequency: After Given the small extension of the evergreen needleleaf forests 16 years, forests affected by more than one fire seldom returned to (original type 1) in the region, we merged these with evergreen broad- their initial forest cover type. leaf forests (original type 2) into one single category of EF, and lastly, We discovered fire-associated forest devastation on a previously we merged all other categories into a new category named “others.” unknown scale. However, forest loss is not subject to purely physi- The MODIS land cover type dataset has annual information, from cal or ecological dynamics in isolation. Instead, synergies among all which annual rasters were extracted for the region and countries of drivers—climate change, land use, increasing drought intensity and Latin America and the Caribbean. frequency, deforestation, and its associated wildfires—can further Combining both datasets from the burnt datasets, we extracted reduce forests areas, compromising carbon reserves (12). Since the all pixels that were forests in 2001. To select only forests affected by extent of forest loss following even one fire event is vast, our research fire after 2001, we removed from the analysis the 544,425 ha of for- emphasizes the importance of committing to an urgent reduction of ests area that were identified as also burnt in 2001. Therefore, the forest fires from any cause to conserve Latin American forests. Given sampled pixels complied with two conditions: (i) They had to have the broader patterns of massive change found in the region, renewed had forests cover in 2001, and (ii) of those forests, only if burnt in regional efforts are urgent to stop irreversible forest change of ep- 2002/2003 were included in our database for analysis. All images ochal planetary consequences. were built using GEE were exported to Google Drive for further processing. We analyzed these datasets using ArcGIS (ESRI) and R statistical MATERIALS AND METHODS software package RStudio with raster-based zonal statistics. We cal- The study region encompasses Latin America, from Mexico in the culated change trajectories defined as sequences of successive con- north to Chile and Argentina in the south, with tropical, subtropical, version of land cover types that provided information on changes and some temperate forests. We included 25 countries. between two or more periods of time. We focus on 2001/2004 changes We used Google Earth Engine (GEE) (42), a cloud-based platform to assess immediate postfire forest cover type changes since our that facilitates access to high-performance computing resources to sampled pixels consisted only of forests burnt in 2002/2003. For the access and collect base information from the MCD64A1 MODIS burnt 2004/2018 period, after forest conversion to other land cover types, area product (43) and MCD12Q1 MODIS yearly land cover type the starting point for analysis does contain all land cover types, in- (32). Both datasets have a spatial resolution of 500 m. MCD64A1 cluding those remaining as forests and we also follow their trajectory has been found to be the most accurate of six different global BA of all over time. To summarize the results, we use alluvial charts and datasets (44–46), is widely used for the regional and global impact include four trajectory classes of stable forest type (when forest type of fire (47), and forms the basis for the Global Fire Atlas (47). Sever- remains the same between periods), stable forest cover (when a spe- al studies have also indicated that this BA product has substantial cific forest type changed to other type but remain as a forest), loss improvements compared to the previous generation of global-scale (when forest cover type is transformed to other nonforest land cover), medium-resolution BA products (48–50). Using the BA information and gain (for cases where forest cover type is recovered). These vi- as a baseline, we built different image collections from 2001 since sualizations where repeated for all forest pixels with frequency of this is the first year of MCD64A1 burned data available using date two or three and more fires over the period of study. filters and reclassifications to reduce the size of the collections and The political administrative limits provided the spatial bases for in preparation for spatial operations. We calculated the fire frequency the extraction and quantification of yearly datasets per forest type, Armenteras et al., Sci. Adv. 2021; 7 : eabd3357 13 August 2021 6 of 8 SCIENCE ADVANCES | RESEARCH ARTICLE fire frequency, and region or country. Only countries with more than 11. B. L. De Faria, P. M. Brando, M. N. Macedo, P. K. Panday, B. S. Soares-Filho, M. T. Coe, 5000 ha of forests affected by fires in 2002/2003 were included in Current and future patterns of fire-induced forest degradation in Amazonia. Environ. Res. Lett. 12, 095005 (2017). country analyses. 12. P. M. Brando, L. Paolucci, C. C. Ummenhofer, E. M. Ordway, H. Hartmann, M. E. Cattau, We adopted a hierarchical Bayesian approach to estimate changes L. Rattis, V. Medjibe, M. T. Coe, J. Balch, Droughts, wildfires, and forest carbon cycling: from various forest cover types while accounting for country-specific A pantropical synthesis. Annu. Rev. Earth Planet. Sci. 47, 555–581 (2019). variation in both the total pixels and the proportion of forest types 13. B. Buma, Disturbance interactions: Characterization, prediction, and the potential burnt. These models were applied for the single fire dataset, with for cascading effects. Ecosphere 6, 1–15 (2015). 14. A. Baccini, W. Walker, L. Carvalho, M. Farina, D. Sulla-Menashe, R. A. Houghton, Tropical burned pixel counts as the response variable and the starting forest forests are a net carbon source based on aboveground measurements of gain and loss. type, final cover type, and country as categorical explanatory variables. Science 358, 230–234 (2017). We generated a series of Poisson-distributed models with increasingly 15. R. J. Keenan, G. A. Reams, F. Achard, J. V. de Freitas, A. Grainger, E. Lindquist, Dynamics complex hierarchical structures using the R package MCMCglmm of global forest area: Results from the FAO Global Forest Resources Assessment 2015. For. Ecol. Manage. 352, 9–20 (2015). (53). In a hierarchical model, the variance explained by the pre- 16. E. Da Ponte, M. Fleckenstein, P. Leinenkugel, A. Parker, N. Oppelt, C. Kuenzer, Tropical dictors can be modeled as applying to the entire sample (often called forest cover dynamics for Latin America using Earth observation data: A review covering “fixed”) or only to certain groups [often called “random” (54)]. Be- the continental, regional, and local scale. Int. J. Remote Sens. 36, 3196–3242 (2015). sides the advantages of accounting for overdispersion in real data 17. M. C. Hansen, P. V. Potapov, R. Moore, M. Hancher, S. A. Turubanova, A. Tyukavina, relative to the Poisson distribution and being able to estimate group-­ D. Thau, S. V. Stehman, S. J. Goetz, T. R. Loveland, A. Kommareddy, A. Egorov, L. Chini, C. O. Justice, J. R. G. Townshend, High-resolution global maps of 21st-century forest cover specific coefficients (e.g., for country effects), the Bayesian approach change. Science 342, 850–853 (2013). also allowed for different levels of variation and covariation in group-­ 18. D. Armenteras, L. Schneider, L. M. Dávalos, Fires in protected areas reveal unforeseen specific effects (e.g., EF is present in all countries, but DB and MF costs of Colombian peace. Nat. Ecol. Evol. 3, 20–23 (2019). are not and so the latter two categories have greater variance). Thus, 19. E. L. Bullock, C. E. Woodcock, C. Souza, P. Olofsson, Satellite-based estimates reveal widespread forest degradation in the Amazon. Glob. Chang. Biol. 26, 2956–2969 (2020). the simplest model had three categorical predictors as sample-wide 20. M. Kalamandeen, E. Gloor, E. Mitchard, D. Quincey, G. Ziv, D. Spracklen, B. Spracklen, regressors and no group-specific coefficients, while the most com- M. Adami, L. E. O. C. Aragão, D. Galbraith, Pervasive rise of small-scale deforestation plex model included country-specific effects allowed to vary de- in Amazonia. Sci. Rep. 8, 1600 (2018). pending on the starting forest type and covary between pairs of 21. D. Armenteras, P. Negret, L. F. Melgarejo, T. M. Lakes, M. C. Londoño, J. García, T. Krueger, starting forest types to different degrees. Depending on model com- M. Baumann, L. M. Davalos, Curb land grabbing to save the Amazon. Nat. Ecol. Evol. 3, 1497–1497 (2019). plexity, models ran for 100,000 to 2 million iterations sampling every 22. E. A. Davidson, A. C. de Araújo, P. Artaxo, J. K. Balch, I. F. Brown, M. M. C. Bustamante, 100 to 1000 iterations in the chain and discarding the first half of all M. T. Coe, R. S. DeFries, M. Keller, M. Longo, J. W. Munger, W. Schroeder, B. S. Soares-Filho, iterations as burn-in. The DIC, with lower values corresponding to C. M. Souza Jr., S. C. Wofsy, The Amazon basin in transition. Nature 481, 321–328 (2012). better fit, was used to evaluate model fit to the data. 23. L. E. O. C. Aragão, Y. E. Shimabukuro, The incidence of fire in Amazonian forests with implications for REDD. Science 328, 1275–1278 (2010). 24. P. M. Brando, D. Silvério, L. Maracahipes-Santos, C. Oliveira-Santos, S. R. Levick, M. T. Coe, M. Migliavacca, J. K. Balch, M. N. Macedo, D. C. Nepstad, L. Maracahipes, E. Davidson, SUPPLEMENTARY MATERIALS G. Asner, O. Kolle, S. Trumbore, Prolonged tropical forest degradation Supplementary material for this article is available at http://advances.sciencemag.org/cgi/ due to compounding disturbances: Implications for CO2 and H2O fluxes. Glob. Chang. Biol. content/full/7/33/eabd3357/DC1 25, 2855–2868 (2019). 25. D. Armenteras, T. M. González, J. Retana, Forest fragmentation and edge influence on fire REFERENCES AND NOTES occurrence and intensity under different management types in Amazon forests. Biol. 1. Food and Agriculture Organization of the United Nations (FAO), Global Forest Resources Conserv. 159, 73–79 (2013). Assessment (Food and Agriculture Organization, 2020). 26. H. Escobar, Bolsonaro’s first moves have Brazilian scientists worried. Science 363, 330–330 2. J. A. Foley, R. Defries, G. P. Asner, C. Barford, G. Bonan, S. R. Carpenter, F. S. Chapin, (2019). M. T. Coe, G. C. Daily, H. K. Gibbs, J. H. Helkowski, T. Holloway, E. A. Howard, C. J. Kucharik, 27. D. S. Nogueira, B. S. Marimon, B. H. Marimon-Junior, E. A. Oliveira, P. Morandi, S. M. Reis, C. Monfreda, J. A. Patz, I. C. Prentice, N. Ramankutty, P. K. Snyder, Global consequences F. Elias, E. C. Neves, T. R. Feldpausch, J. Lloyd, O. L. Phillips, Impacts of fire on forest of land use. Science 309, 570–574 (2005). biomass dynamics at the southern amazon edge. Environ. Conserv. 46, 285–292 (2019). 3. F. Seymour, N. L. Harris, Reducing tropical deforestation. Science 365, 756–757 (2019). 28. M. Hirota, M. Holmgren, E. H. Van Nes, M. Scheffer, Global resilience of tropical forest 4. P. G. Curtis, C. M. Slay, N. L. Harris, A. Tyukavina, M. C. Hansen, Classifying drivers of global and savanna to critical transitions. Science 334, 232–235 (2011). forest loss. Science 361, 1108–1111 (2018). 29. Z. Liu, A. P. Ballantyne, L. A. Cooper, Biophysical feedback of global forest fires on surface 5. N. Andela, D. C. Morton, L. Giglio, Y. Chen, G. R. van der Werf, P. S. Kasibhatla, R. S. DeFries, temperature. Nat. Commun. 10, 214 (2019). G. J. Collatz, S. Hantson, S. Kloster, D. Bachelet, M. Forrest, G. Lasslop, F. Li, S. Mangeon, 30. G. R. van der Werf, J. T. Randerson, L. Giglio, G. J. Collatz, M. Mu, P. S. Kasibhatla, J. R. Melton, C. Yue, J. T. Randerson, A human-driven decline in global burned area. D. C. Morton, R. S. DeFries, Y. Jin, T. T. van Leeuwen, Global fire emissions Science 356, 1356–1362 (2017). and the contribution of deforestation, savanna, forest, agricultural, and peat fires 6. L. E. O. C. Aragão, L. O. Anderson, M. G. Fonseca, T. M. Rosan, L. B. Vedovato, F. H. Wagner, (1997–2009). Atmos. Chem. Phys. Discuss. 10, 16153–16230 (2010). C. V. J. Silva, C. H. L. Silva Junior, E. Arai, A. P. Aguiar, J. Barlow, E. Berenguer, M. N. Deeter, 31. A. A. Alencar, P. M. Brando, G. P. Asner, F. E. Putz, Landscape fragmentation, severe L. G. Domingues, L. Gatti, M. Gloor, Y. Malhi, J. A. Marengo, J. B. Miller, O. L. Phillips, drought, and the new Amazon forest fire regime. Ecol. Appl. 25, 1493–1505 (2015). S. Saatchi, 21st Century drought-related fires counteract the decline of Amazon 32. M. Friedl, D. Sulla-Menashe, MCD12Q1 MODIS/Terra+Aqua Land Cover Type Yearly L3 deforestation carbon emissions. Nat. Commun. 9, 536 (2018). Global 500m SIN Grid V006 [Data set] (NASA EOSDIS Land Processes DAAC, 2019); 7. D. M. J. S. Bowman, J. K. Balch, P. Artaxo, W. J. Bond, J. M. Carlson, M. A. Cochrane, C. M. D’Antonio, https://doi.org/10.5067/MODIS/MCD12Q1.006. R. S. Defries, J. C. Doyle, S. P. Harrison, F. H. Johnston, J. E. Keeley, M. A. Krawchuk, C. A. Kull, 33. C. C. Nóbrega, P. M. Brando, D. V. Silvério, L. Maracahipes, P. de Marco, Effects J. B. Marston, M. A. Moritz, I. C. Prentice, C. I. Roos, A. C. Scott, T. W. Swetnam, of experimental fires on the phylogenetic and functional diversity of woody species G. R. van der Werf, S. J. Pyne, Fire in the earth system. Science 324, 481–484 (2009). in a neotropical forest. For. Ecol. Manage. 450, 117497 (2019). 8. M. A. Cochrane, Tropical Fire Ecology (Springer, 2009). 34. J. Barlow, C. A. Peres, B. O. Lagan, T. Haugaasen, Large tree mortality and the decline 9. N. Hosonuma, M. Herold, V. de Sy, R. S. de Fries, M. Brockhaus, L. Verchot, A. Angelsen, of forest biomass following Amazonian wildfires. Ecol. Lett. 6, 6–8 (2002). E. Romijn, An assessment of deforestation and forest degradation drivers in developing 35. L. Zhou, Y. Wang, Y. Chi, S. Wang, Q. Wang, Contrasting post-fire dynamics between countries. Environ. Res. Lett. 7, 44009 (2012). Africa and South America based on MODIS observations. Remote Sens. 11, 1074 10. Y. Chen, J. T. Randerson, D. C. Morton, R. S. DeFries, G. J. Collatz, P. S. Kasibhatla, L. Giglio, (2019). Y. Jin, M. E. Marlier, Forecasting fire season severity in South America using sea surface 36. P. G. Curtis, C. M. Slay, N. L. Harris, A. Tyukavina, M. C. Hansen, Supplementary Materials: temperature anomalies. Science 334, 787–791 (2011). Classifying drivers of global forest loss. Science 361, 1108–1111 (2018). Armenteras et al., Sci. Adv. 2021; 7 : eabd3357 13 August 2021 7 of 8 SCIENCE ADVANCES | RESEARCH ARTICLE 37. J. W. Veldman, E. Buisson, G. Durigan, G. W. Fernandes, S. le Stradic, G. Mahy, D. Negreiros, 50. J. A. Rodrigues, R. Libonati, A. A. Pereira, J. M. P. Nogueira, F. L. M. Santos, L. F. Peres, G. E. Overbeck, R. G. Veldman, N. P. Zaloumis, F. E. Putz, W. J. Bond, Toward an old-growth A. Santa Rosa, W. Schroeder, J. M. C. Pereira, L. Giglio, I. F. Trigo, A. W. Setzer, How well do concept for grasslands, savannas, and woodlands. Front. Ecol. Environ. 13, 154–162 global burned area products represent fire patterns in the Brazilian Savannas biome? (2015). An accuracy assessment of the MCD64 collections. Int. J. Appl. Earth Obs. Geoinf. 78, 38. A. Tyukavina, M. C. Hansen, P. V. Potapov, S. V. Stehman, K. Smith-Rodriguez, C. Okpa, 318–331 (2019). R. Aguilar, Types and rates of forest disturbance in Brazilian Legal Amazon, 2000–2013. 51. T. Fanin, G. R. Van Der Werf, Relationships between burned area, forest cover loss, Sci. Adv. 3, e1601047 (2017). and land cover change in the Brazilian Amazon based on satellite data. Biogeosciences 12, 39. M. A. Cochrane, Fire science for rainforests. Nature 421, 913–919 (2003). 6033–6043 (2015). 40. M. Uriarte, N. Schwartz, J. S. Powers, E. Marín-Spiotta, W. Liao, L. K. Werden, Impacts 52. A. N. Safronov, Effects of climatic warming and wildfires on recent vegetation changes of climate variability on tree demography in second growth tropical forests: in the Lake Baikal basin. Climate 8, 57 (2020). The importance of regional context for predicting successional trajectories. Biotropica 48, 53. J. D. Hadfield, MCMC methods for multi-response generalized linear mixed models: 780–797 (2016). The MCMCglmm R package. J. Stat. Softw. 33, 29555 (2010). 41. M. A. Cochrane, Synergistic interactions between habitat fragmentation and fire 54. A. Gelman, J. Hill, Data Analysis using Regression and Multilevel/Hierarchical Models in evergreen tropical forests. Conserv. Biol. 15, 1515–1521 (2001). (Cambridge Univ. Press, 2017). 42. N. Gorelick, M. Hancher, M. Dixon, S. Ilyushchenko, D. Thau, R. Moore, Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sens. Environ. 202, Acknowledgments 18–27 (2017). Funding: NAS Subaward Letters nos. 2000007526 and 2000010972 awarded to D.A. The 43. Giglio, L., C. Justice, L. Boschetti, D. Roy, MCD64A1 MODIS/Terra+Aqua Burned Area Chilean National of Scientific and Technological Research (CONICYT) with Project Formación Monthly L3 Global 500m SIN Grid V006 [Data set] (NASA EOSDIS Land Processes DAAC, de Redes Internacionales para Investigadores en Etapa Inicial N° REDI170329 awarded to 2015). A.H.-M. L.M.D. was supported in part by NSF-DGE 1633299. Author contributions: D.A. 44. M. Padilla, S. V. Stehman, R. Ramo, D. Corti, S. Hantson, P. Oliva, I. Alonso-Canas, designed research; D.A., A.M., and J.S.B. performed research and data analysis; L.M.D. A. V. Bradley, K. Tansey, B. Mota, J. M. Pereira, E. Chuvieco, Comparing the accuracies conducted visualizations and Bayesian models; D.A., A.M., J.S.B., A.H.-M., C.Z.-E., T.M.G.-D., of remote sensing global burned area products using stratified random sampling M.C.M, L.M.D., and J.R. analyzed the results; and D.A. wrote the paper with inputs from and estimation. Remote Sens. Environ. 160, 114–121 (2015). T.M.G-D., L.M.D., and J.R. All authors revised and approved the final text. Competing interests: 45. D. Fornacca, G. Ren, W. Xiao, Performance of three MODIS fire products (MCD45A1, The authors declare that they have no competing interests. Data and materials availability: MCD64A1, MCD14ML), and ESA Fire_CCI in a mountainous area of Northwest Yunnan, All data needed to evaluate the conclusions in the paper are present in the paper and/or the China, characterized by frequent small fires. Remote Sens. (Basel) 9, 1131 (2017). Supplementary Materials. R codes for Alluvial Graph figures and Bayesian models are available 46. N. Quintero, O. Viedma, I. R. Urbieta, J. M. Moreno, Assessing landscape fire hazard by at Dryad, Dataset, https://doi.org/10.5061/dryad.9cnp5hqh6. multitemporal automatic classification of landsat time series using the Google Earth Engine in West-Central Spain. Forests 10, 518 (2019). Submitted 16 June 2020 47. N. Andela, D. C. Morton, L. Giglio, R. Paugman, Y. Chen, S. Hantson, G. R. van der Werf, Accepted 24 June 2021 J. T. Randerson, The Global Fire Atlas of individual fire size, duration, speed, and direction. Published 13 August 2021 Earth Syst. Sci. Data 11, 529–552 (2019). 10.1126/sciadv.abd3357 48. L. Giglio, L. Boschetti, D. P. Roy, M. L. Humber, C. O. Justice, The Collection 6 MODIS burned area mapping algorithm and product. Remote Sens. Environ. 217, 72–85 (2018). Citation: D. Armenteras, L. M. Dávalos, J. S. Barreto, A. Miranda, A. Hernández-Moreno, C. Zamorano-Elgueta, 49. M. L. Humber, L. Boschetti, L. Giglio, C. O. Justice, Spatial and temporal intercomparison T. M. González-Delgado, M. C. Meza-Elizalde, J. Retana, Fire-induced loss of the world’s most of four global burned area products. Int. J. Digit. Earth. 12, 460–484 (2018). biodiverse forests in Latin America. Sci. Adv. 7, eabd3357 (2021). Armenteras et al., Sci. Adv. 2021; 7 : eabd3357 13 August 2021 8 of 8