Forest recovery pathways after fire, drought, and windstorms in southeastern Amazonia

Significance

As climate and land-use change intensify in Amazonia, compound disturbances could push southeastern Amazon forests toward large-scale collapse, including transitions to alternative states. Using a 20-y experimental fire study in southeastern Amazonia, we evaluated how diversity and functional traits responded to repeated fires, edge effects, droughts, and windstorms. When fires ceased, recovery occurred rapidly, with exception along edges adjacent to open agriculture areas. We observed changes in communities toward drought-tolerant, generalist species; Amazonian specialists continued to decline, but savanna woody species rarely established. Grass cover along edges also diminished, leaving mostly shade-tolerant grasses. These results show that heavily degraded, grass-invaded forests can remain resilient—driven largely by the recovery capacity of generalist species, even with reduced tree diversity.

Abstract

The future of tropical forests depends on their ability to resist and recover from multiple disturbances. Here, we evaluated how edge effects, experimental fires, extreme droughts, and blowdowns reshaped forest structure, composition, and functional traits over two decades in the Amazon–Cerrado transition. Initially, forests resisted low-intensity fires, but subsequent high-intensity fires during severe droughts sharply increased susceptibility to further disturbances. Along forest edges bordering agriculture, these compound disturbances drove losses of tree species richness, declines in Amazonian forest-specialist species, and shifts toward generalists with broad distributions, indicating increased compositional homogenization (i.e., reduced taxonomic diversity and dominance of generalist species). Once fires ceased, recovery trajectories diverged: Interior forests rapidly regained woody species richness and composition, whereas edge forests recovered more slowly. During postfire recovery, embolism-resistant species (lower P50 values) became more common, yet communities exhibited lower hydraulic safety margins (increased vulnerability to drought). At the same time, generalist species remained abundant, forest-specialist declined, and only a single woody species typical of savanna was established. Although grasses initially colonized fire-altered edges—especially light-demanding exotic Andropogon gayanus—they declined once fires stopped, leaving only small patches of shade-tolerant C3 species. Consequently, we find little evidence that fire-degraded forests transition toward persistent savanna, although recurrent fires or future climatic changes could drive long-lasting degradation or human-derived savannas. Together, these results show that even highly degraded, grass-invaded forests can recover in the absence of new disturbances, but the communities that reassemble remain vulnerable to renewed fire, drought, or windthrow.

Synergistic interactions between natural and anthropogenic disturbances are fundamentally altering the structure, composition, and functioning of tropical forests. In southeastern Amazonia, the co-occurrence of fires, droughts, and fragmentation has already caused widespread degradation and may accelerate large-scale transitions to nonforest ecosystems in the coming decades through interactions between land use and climate change (1–4). Addressing the “Amazon dieback hypothesis” requires quantifying both forest resistance—the probability that trees survive chronic stressors—and resilience—the capacity of forests to recover structure, composition, and diversity through natural regeneration (5, 6). Yet, while functional traits often explain species performance under single stressors, their predictive power weakens when multiple disturbances interact synergistically, given a trait that enhances survival under one stressor may reduce it under another (7). This makes it unclear whether tree assemblages will follow predictable directional changes toward communities more resistant to future disturbances or will fundamentally change over time.

A central ecological question is whether such disturbances restructure Amazonian tree communities through the loss of disturbance-sensitive taxa, the establishment of species adapted to novel conditions, or both. Functional traits provide important insights into these dynamics. Bark thickness (BT) and wood density (SSD) predict resistance to wildfires (8, 9); hydraulic traits such as P₅₀ (the water potential at which 50% of hydraulic conductivity is lost) and hydraulic safety margin (HSM) indicate drought vulnerability (10–12); tree size and wood density influence drought-, windthrow-, and fire-related mortality (13, 14); and specific leaf area (SLA) is linked to photosynthetic efficiency and postdisturbance recovery potential (15). Shifts in community-weighted mean (CWM) trait values therefore can signal directional changes in the capacity of plant communities to recover from and cope with recurrent or novel disturbances. However, while these traits capture species responses to individual stressors, their predictive power under compound disturbances may be limited--leaving uncertainty regarding whether assemblages reorganize in consistent ways toward more resistant communities (4).

Climatic niches and habitat preference (16, 17) offer complementary metrics to evaluate directional change in forest trajectories. For instance, tree species with distribution concentrated in the humid Amazon may be especially vulnerable to droughts and dry-hot edges bordering agricultural fields (18, 19), while taxa with drier niches or savanna-affiliated tree species would be expected to show traits conferring greater fire and drought resistance (e.g., thick bark, high wood density) (8, 20). However, it remains unclear whether biogeographical patterns reliably predict directional changes in communities impacted by disturbances. If climatic niche and habitat preference are correlated with susceptibility, we would expect disturbances along the drier Amazon margins to favor generalist species (e.g., species common to different biomes) and ultimately Cerrado species at the expense of humid-forest specialists. In the ecotone with the Cerrado, the likelihood of cross-biome species may increase, with savanna-affiliated species potentially becoming more abundant in forests impacted by disturbances.

Disturbances can also trigger positive fire–vegetation feedbacks that amplify forest vulnerability to abrupt ecosystem transitions (3, 21, 22). Canopy loss is hypothesized to promote grass invasion along forest edges neighboring agricultural fields, which could suppress tree regeneration, increase flammability, and raise the probability of recurrent fires (23). If fire ignition sources are present, this cycle can lead to severe forest degradation and even to ecosystem transition into human-derived savannas (3, 24, 25)—or degraded grassy systems dominated by exotic and Cerrado-native species, as shown in experimental forest in southeastern Amazonia (SI Appendix, Fig. S1) and in fire-impacted areas of the Xingu Indigenous Territory (26).

Here, we analyze the long-term dynamics of experimental forests in southeastern Amazonia subjected to experimental fires (SI Appendix, Fig. S1). Our study site lies within 20 km of the forest–savanna ecotone, where cross-biome species pools increase the potential for major shifts in composition. The experimental area includes three 50-ha plots exposed to annual burns (B1yr), triennial burns (B3yr), or unburned controls from 2004 to 2010 (27). These plots also experienced natural disturbances, including blowdown events (2012 and 2019) and severe droughts (2007, 2010, and 2015–2016) (28–30). All experimental forest plots have edges that were adjacent to pasture from 1981 to 2007, followed by soybean cultivation after 2007 (25, 31). Grass invasion along the forest edges of the experimental area led to high-intensity and more severe fires, with slow recovery rates once the experimental fires ended (25, 27, 32).

We evaluate trajectories of species composition, diversity, functional traits, and precipitation niche metrics to test two alternative recovery pathways following the cessation of burning in 2010. One involves the rise of acquisitive strategies [e.g., fast growth and rapid resource uptake (33, 34)], disturbance-tolerant generalists with broad climatic niches and declining grass cover. The other reflects recruitment of conservative [e.g., slow growth and durable tissues and resource storage (33, 34)], stress resistant savanna-affiliated tree species, potentially accompanied by increased grass cover and reduced diversity. We predict that the impacts of multiple disturbances will be most pronounced at forest edges, where large-tree mortality from fire, drought, and blowdowns has been consistently higher than in forest interiors.

Results

Species Diversity and Richness.

The experimental fires and subsequent disturbances drove substantial reductions in forest biomass, forest height, and stand basal area (see (27, 30, 31); Fig. 1 and SI Appendix, Fig. S1). Despite these changes in forest structure, tree species richness remained relatively constant during the study period in the forest interior. For instance, species richness in the forest interior of plots burned annually (B1yr) and triennially (B3yr) varied from 71 and 73 to 69 and 65 species, respectively, between 2004 to 2024, while in the Control it increased by 21% (from 62 to 75 species). In contrast, we observed clear reductions in species richness along the forest edges (0 to 200 m into the forest) of the burned plots (Fig. 1 and SI Appendix, Figs. S2 and S3). From 2004 to 2011, species richness at the forest edge fell by 20.3% in B1yr (from 64 to 51 species) and by 46.2% in B3yr (from 65 to 35 species). Species richness at plot edges continued to decline after burning ended, reaching a minimum in 2014. Although species richness at the forest edge started to recover after 2016, it was still 31.3% and 50.8% lower in B1yr and B3yr, respectively, compared to prefire values by 2024.

Fig. 1.

Temporal changes in species richness (A), total density of individuals (B), and basal area (C) measured along the forest edge (0 to 200 m into the forest; Left panels) and the forest interior (>200 m inside the forest; Right panels) for individuals with ≥10 cm DBH in three experimental plots [burned every 3 y (B3yr), annually (B1yr), and unburned (Control)] in southeastern Amazonia at Tanguro Research Station (Mato Grosso, Brazil). The last fire event was in 2010 and the consequent beginning of forest recovery (2010 to 2024). The dashed lines represent the phases of the experiment: Phase I, low fire impact (gray line); Phase II, more severe compounded disturbances (blue line); Phase III, forest recovery.

Species Composition.

Beyond changes in species richness along the forest edges, we observed pronounced shifts in species composition associated with edge effects and the experimental fires for both large and small trees (Fig. 2 and SI Appendix, Figs. S4 and S5). Nonmetric multidimensional scaling analysis (NMDS) revealed strong temporal changes in tree community composition that differed among treatments (SI Appendix, Fig. S5). Between 2004 and 2007, tree species composition across the three treatment plots showed high floristic similarity, occurring in close proximity in ordination space. However, this pattern began to diverge following the last prescribed fires in 2010, with a clear directional shift from the early period evident by 2014 (SI Appendix, Fig. S5). At the most recent census in 2024, tree communities growing in the burned plots were still clearly distinct from both the Control and the earlier periods (Fig. 2 and SI Appendix, Fig. S5). These compositional shifts were predominantly observed at the forest edge, with less pronounced changes in the forest interior (Fig. 2).

Fig. 2.

Change in woody species composition for individuals with ≥5 cm DBH in three experimental plots (B3yr, B1yr, and Control) in southeastern Amazonia (Tanguro Research Station, Mato Grosso, Brazil), represented by NMDS. Upper panel show changes in species composition for the forest edge (A) and Lower panel show changes in species composition for the forest interior (B).

The observed shift in species composition was driven by a decline in late successional species during postdisturbance recovery, accompanied by an increase in the abundance of early successional species (Fig. 3 and SI Appendix, Figs. S6–S8). For instance, the proportion of pioneer species more than tripled in the burned plots (from 17.1 to 59% in B1yr; and from 21.6 to 72.6% in B3yr; SI Appendix, Fig. S6), while it dropped from 21.6 to 20.4% in the Control from 2004 to 2024. Along the forest edges, the proportion of pioneer species reached 80.3% and 93.3% in B1yr and B3yr, respectively (SI Appendix, Fig. S7). Additionally, most species that declined in abundance within the burned plots also showed reductions in the Control plots—11 of the 18 species (Fig. 3 and SI Appendix, Figs. S8 and S9). In the Control, this shift was more noticeable after 2012. Although natural disturbances affected all plots [e.g., blowdowns in 2012 (28) and 2019, and isolated treefall gaps across years], they did not lead to a large increase in pioneer species as in the burned plots. Despite their proximity to savanna sites, the burned plots showed no influx of woody savanna species over time (SI Appendix, Table S2).

Fig. 3.

Difference in relative density values of the families between Prefire (2004) and Recovery (2024) periods for individuals with ≥ 10 cm DBH in three experimental plots (B3yr, B1yr, and Control) in southeastern Amazonia at Tanguro Research Station (Mato Grosso, Brazil). Red indicates a reduction in the species relative density between prefire and recovery periods, and light blue indicates an increase. The species are classified according to their respective successional groups [late successional (dark blue) and pioneer (green)]. Only species with absolute abundance values greater than or equal to 10 individuals in at least one of the treatments are presented in at least one of the treatments. All scientific names corresponding to the codes presented in this figure are available in SI Appendix, Table S2.

Forest-Grass Competition.

After the high-severity fires of 2007, grasses began to invade the edges of the experimental forest, initially dominated by the native Cerrado species Aristida longifolia (25). With subsequent fires, this grass expansion intensified as canopy cover decreased, reaching a maximum in 2012—2 y after the last fire in 2010—when most of the edges were colonized by an African invasive pasture grass Andropogon gayanus, widely planted in the region (Fig. 7). By 2015, grass cover remained comparable to 2012, in part due to delayed mortality of large trees that created light conditions favorable for A. gayanus (30) and other native and exotic grasses. However, the recovery of tree species from 2016 onward promoted rapid canopy closure, sharply reducing grass cover along the forest edges of the burned plots (Fig. 4). By 2025, grass cover had declined to 12% at forest edges of the burned plots, although most of this was due to the presence of A. longifolia in the forest understory. In contrast, grass cover in control edges consistently remained below 1% throughout the monitoring period.

Fig. 7.

(A) Distribution of occurrence points along the Amazon and Cerrado of the 120 woody species sampled in southeastern Amazonia at Tanguro Research Station (Mato Grosso, Brazil). (B) Map of the MCWD of species measured along in the Amazon and the Cerrado. (C) Temporal changes in MCWD and (D) habitat preference of species measured along the forest edge and the forest interior in three experimental plots (B3yr, B1yr, and Control) (E and F). MCWD values close to zero represent species with a precipitation niche more adapted to humid regions. The last fire event was in 2010 and the consequent beginning of forest recovery (2010 to 2024). The dashed lines represent the phases of the experiment: Phase I, low fire impact (gray line); Phase II, more severe compounded disturbances (blue line); Phase III, forest recovery. The colored bands represent bootstrapped CI.

Fig. 4.

(A) Land-cover dynamics in plots burned annually (B1yr) and triennial (B3yr) for 2012, 2015, 2022, and 2025. (B) Map of grass dynamics for 2012, 2015, 2022, and 2025 in three experimental plots (B3yr, B1yr, and Control).

Drought-Resistant Species Performance.

Our results indicate that changes in forest diversity substantially altered community-weighted traits linked to tree vulnerability to drought. Based on P50 and leaf water potential values in the burned plots measured at the peak of the 2021 dry season, we observed an increase in the abundance of tree species with greater embolism resistance (more negative P50 values) and more negative leaf water potentials at both predawn and midday (Fig. 5 and SI Appendix, Fig. S10). This trend was especially pronounced along the forest edge of the burned plots (SI Appendix, Figs. S11 and S12). These findings suggest that fire-affected tree communities have become more embolism-resistant. On the other hand, assuming the 2021 measurements of water potentials are representative of conditions in prior years, the more negative P50s likely offer reduced buffering capacity against soil and atmospheric drying, as evidenced by lower HSMs (HSM₅₀) in the burned plots compared to the control plot in 2021 (Fig. 5 and SI Appendix, Fig. S11). This decline in HSM became more apparent with the recruitment of pioneer species with inherently lower HSM values along the edges of burned plots (SI Appendix, Fig. S13 and S14). Yet, CWM values for HSM were substantially above zero across all treatment plots.

Fig. 5.

Temporal changes in the values of water stress-related functional traits weighted by abundance (CWM) for individuals with ≥10 cm DBH of tree communities of the three experimental plots (B3yr, B1yr, and Control) in southeastern Amazonia at Tanguro Research Station (Mato Grosso, Brazil). (A) In site dry season leaf water potential at the predawn (Ψ_{predawn [MPa]}) and (B) midday (Ψ_{midday [MPa]}). (C) Xylem water potential at which 50% of the conductance is lost (P_{50 [MPa]}) and (D) HSM (HSM_{50 [MPa]}). The last fire event was in 2010 and the consequent beginning of forest recovery (2010 to 2024). The dashed lines represent the phases of the experiment: Phase I, low fire impact (gray line); Phase II, more severe compounded disturbances (blue line); Phase III, forest recovery. The colored bands represent bootstrapped CI.

Resource Use and Fire Resistance Affect Species Demography.

In addition to changes in hydraulic traits, we observed community-level changes in several other functional traits including plant height, wood density, bark thickness, and SLA (Fig. 6 and SI Appendix, Fig. S15). Between 2004 and 2024, dominant tree species in the burned plots were shorter (22.5% for B1yr and 28% for B3yr), had slightly lower wood density (reductions of 5% for B1yr and 4% for B3yr), and had thinner leaves with high SLA (increase of 9.5% for B1yr and 4.6% for B3yr). We also observed changes in absolute bark thickness compared to prefire values. During the burning period, bark thickness increased in the burned plots compared to the Control, likely driven by the selective mortality of thin-barked trees. Subsequently, with the recruitment of fast-growing, thin-barked species, we observed comparable bark thickness values between the burned plots and the Control. All these changes were more pronounced at the forest edge than in the interior (SI Appendix, Figs. S16 and S17). In the Control, we observed an increase in SLA, while bark thickness, plant height, and wood density did not change substantially over time. All observed changes in the burned plots were more related to the recruitment of new individuals than to mortality (SI Appendix, Figs. S18 and S19).

Fig. 6.

Temporal change in the scaled values of fire resistance functional traits weighted by abundance (CWM) for individuals with ≥10 cm DBH of tree communities of the three experimental plots (B3yr, B1yr, and Control) in southeastern Amazonia at Tanguro Research Station (Mato Grosso, Brazil). The last fire event was in 2010 and the consequent beginning of forest recovery (2010 to 2024). The dashed lines represent the phases of the experiment: Phase I, low fire impact (gray line); Phase II, more severe compounded disturbances (blue line); Phase III, forest recovery. (A) BT: bark thickness (mm); (B) Hmax: maximum plant height (m); (C) SSD: stem-specific density (mg mm⁻³); (D) SLA: specific leaf area (cm² g⁻¹). The colored bands represent bootstrapped CI.

Species Precipitation Niche and Forest Recovery.

To assess how changes in species composition and abundance were linked to species’ precipitation niches (SI Appendix, Fig. S20) and habitat preference (SI Appendix, Fig. S21), we calculated the weighted averages of habitat preference (17) and maximum climatological water deficit (MCWD) (16) based on the relative abundance of each tree species across Amazonia and Cerrado. The habitat preference values ranged from 0 (savanna-specialists) to 1 (forest-specialists), with intermediate values (0.3 to 0.7) representing generalist species occurring in both habitats (forest and savanna); the weighted average MCWD represented the fundamental climatic conditions across the region—including the dry-season intensity—of our sampled species.

If changes in species composition were determined by the species climate niches, we predicted that multiple disturbances could drive a shift toward a forest tree community dominated by generalist (i.e., species commonly found in savannas and forests) and savanna-specialist species (i.e., species predominantly found in savannas), rather than forest-specialists (i.e., species commonly found in forest habitats). Moreover, we would expect establishment of tree species with precipitation niches associated with drier and hotter (i.e., savanna-specialist) conditions. Supporting these expectations, the community-weighted habitat preference declined in burned plots due to the loss of forest-specialist species (Fig. 7D and SI Appendix, Fig. S22). We also observed a shift toward tree assemblages representing drier and hotter climate niches, accompanying drops of 36.6% (B1yr) and 41.1% (B3yr) in the MCWD from 2004 to 2024 (Fig. 7C and SI Appendix, Fig. S22). In burned plots, this drier climatic niche was associated with increase in the abundance of generalist species (i.e., occurring in savannas and forest; SI Appendix, Fig. S23). In contrast, changes in the habitat preference and precipitation niche remained relatively constant in the Control plot (Fig. 7 and SI Appendix, Fig. S22).

The observed shifts in habitat preference, and climatic niches were even more pronounced at the forest edge. For instance, the habitat preference dropped from 0.84 to 0.74 in B1yr and 0.84 to 0.73 in B3yr between 2004 and 2024 (Fig. 7d). Both declines were driven largely by an increase in the relative abundance of several generalist species, including Mabea fistulifera, Tachigali vulgaris, and Casearia grandiflora, along with a slight decrease in forest-specialist species in both the B1yr and B3yr plots (Figs. 3 and 7 and SI Appendix, Fig. S23). These shifts in species abundance further drove community-weighted niche metrics toward values representing drier and hotter conditions, with reductions of 59.1% in B1yr and 59.4% in B3yr for MCWD from 2004 to 2024 (Fig. 7 and SI Appendix, Figs. S26–S28).

Discussion

This study examined how multiple disturbances reshape Amazonian tree diversity, composition, functional traits, and susceptibility to future disturbance events, while assessing whether species distributions help explain directional trajectories. We identified three phases of forest response to those disturbances. In the first phase (2004 to 2007), prescribed fires under normal climatic conditions had no clear effects on species richness, functional diversity, or composition despite structural changes (30, 31). These results support the hypothesis that tropical forests are apparently resilient to low-intensity disturbances (4, 30, 31, 35–37). The second phase (2007 to 2010) was characterized by major mortality spikes in the years following the high-intensity fires of 2007 and 2010 (27, 30). These effects were strongest along forest edges, compared to interiors, resulting in significant reductions in species richness and shifts in composition. However, functional traits remained relatively stable, likely facilitated by the presence of species with diverse functional traits that provide multiple strategies for coping with disturbances (6). The third phase (2011 to 2024)—following the end of experimental burns—was characterized by contrasting recovery trajectories: The forest interior rapidly regained predisturbance diversity, whereas edges recovered more slowly, likely because blowdown events and experimental fires were much more severe in these areas (27, 28, 30).

Generalization or Savannization?

The directional shifts in species composition following high-intensity fires were striking, especially along forest edges. These shifts were primarily driven by increased mortality of woody forest-specialist species and an increase of generalist ones with broad geographical distributions, resulting in a decrease in both the species habitat preference and the MCWD (Figs. 3 and 7). Although our experimental forest is located within a forested region of the Amazon Forest-Cerrado ecotone—but about 20 km from open savannas and 5 km from woodland savannas (“Cerradão”)—its recovery was not marked by an influx of woody savanna-specialist species, as hypothesized by previous studies (3, 24, 25, 38, 39). Instead, we observed increased dominance of woody generalist species commonly found across both the forest and savanna formations. This observation provides elements that contradict the hypothesis that woody savanna species would preferentially colonize our fire-degraded experimental forest (23). Alternatively, it suggests that compounding (e.g., fire, severe drought, and blowdowns) and repeated disturbances suppressed the dominance of forest-affiliated and late-successional species, favoring instead generalist and early successional (fast-growing) ones. Similar dynamics have been observed in forests impacted by disturbances, leading to a “homogenization” (e.g., reduced taxonomic diversity and dominance of generalist species) in these environments (40–42). In savanna environments, the absence of fire in mesic areas has even led to the colonization by woody species due to their high dispersal capacity (43, 44).

Nevertheless, we observed a transient invasion of opportunistic Cerrado-native (e.g., A. longifolia and Imperata sp.) and African-pasture grasses in the same experimental area following high-intensity fires and a subsequent blowdown event (25, 28, 30). This highlights the potential role of herbaceous savanna species invasion in altering forest dynamics and accelerating degradation when fire is present. Had experimental burning continued and ignition sources been recurrent, grasses likely would have expanded across the experimental area, leading to a potential long-term transition from forest to a human-derived savanna (23). Similar patterns may be occurring in the Xingu Indigenous Territory, where fire management practices and the abundance of grass propagules have promoted similar feedback dynamics (26). Despite the initial inhibition of forest regeneration by grasses along the forest edges, in the absence of new ignitions, grasses declined—particularly light-demanding species such as A. gayanus—while more shade-tolerant C3 species (A. longifolia) became restricted to small patches (Fig. 4). Thus, we speculate that it is unlikely that in the absence of fire that a savanna-like environment would persist over time with no fundamental changes in climate. Overall, these results highlight that effective fire management preventing escaped burns can allow highly degraded forests with presence of grasses to recover—a necessity that will only grow as fires become more frequent under future climate scenarios (26, 45).

Functional Traits and Future Susceptibility.

The observed changes in species composition following experimental fires raise questions about whether burned forests become more or less vulnerable to subsequent disturbances. Theoretically, fires can act as a strong selective filter that decreases the abundance of species most vulnerable to burning (e.g., thinner-barked species), potentially enhancing the community’s overall fire resistance (8, 9). This filter could influence regeneration patterns by increasing the seed rain of fire-resistant species, driving a gradual shift toward tree communities that are better able to withstand fires over the long term (29). Our findings suggest that experimental fires do act as a selective environmental filter, but do not consistently enhance resistance to fire disturbance (Fig. 6). While there was an increase in CWM for BT during the burning period, we observed a subsequent decrease during the postfire period, which likely increased forest vulnerability to future fires. This pattern can be attributed to regeneration dynamics in fire-disturbed areas, where colonization by thinner-barked pioneer species constrains the development of fire-resistant communities over the short term (34). Furthermore, this context-dependent result could stem from the combined disturbances. While some functional traits enhance resistance to specific disturbances, they may simultaneously diminish resistance against others. For example, large trees (typically with thick bark) tend to be more fire-resistant but also more susceptible to blowdowns and droughts (46, 47).

Given the occurrence of multiple droughts and the influx of hot, dry air from neighboring agricultural fields, tree communities in our experimental forests may be shifting toward greater drought resistance. This is indicated by the recruitment of generalist species along forest edges in the burned areas, which drove an increase in the community’s average embolism resistance (P₅₀ values)— suggesting that trees colonizing the burned areas are more drought resistant (Figs. 5 and 7). However, despite the increase in embolism resistance, community-level drought vulnerability remained similar to the Control, suggesting greater exposure to drought conditions, potentially due to reduced access to deep soil water and/or higher vapor pressure deficit (VPD) in burned plots. Such changes would elevate overall water stress, thereby increasing forest vulnerability and susceptibility to extreme drought events. Our analysis indicates that while these experimental forests in southeastern Amazonia show early signs of higher vessel resistance to drought, the postdisturbance tree communities could begin to operate at critical thresholds of ecophysiological tolerance in the face of land use and climate changes (48). If disturbances keep increasing in severity and frequency and trees fail to adapt, these forests risk losing the long-term stability and functional integrity of the ecosystem, given that some species may have less access to deep soil moistures. Interestingly, although the experimental fires accelerated changes in community structure and hydraulic traits, we observed similar—but less pronounced—changes in the unburned Control plots. This suggests that broader regional changes in climate may be an underlying stressor to these highly dynamic forests (49).

Conclusions

Our study sheds light on these forest dynamics in southern Amazonia, a region experiencing rapid changes, underscoring the compounding effects of fires, drought, and blowdowns on the structure, composition, and functioning of these tropical forests. Even in a highly degraded state, these ecosystems exhibit remarkable resilience, marked by an influx of generalist species with acquisitive strategies (33). The removal of the most vulnerable species by disturbances has led to the establishment of new species that display greater resistance. In this context, a pivotal question arises: Can these degraded forests in southeastern Amazonia transition to an alternative open stable state upon recovery? Our findings suggest that with the suppression of prescribed fires, the forest interior rapidly recovers its structure and functioning. However, the forest recovery along the edge takes much longer based on our metrics of diversity and functional traits. Importantly, our results also highlight that communities in early successional states could be more susceptible to subsequent disturbance events, including fires, severe droughts, and blowdowns.

Despite postfire recovery of some ecosystem services (e.g., carbon and water flows) with the growth of new recruits (30), diversity has not reached predisturbance levels and species composition remained different, suggesting that degraded areas in southeastern Amazonia are resilient but will likely require much longer than 14 y to recover all forest properties (e.g., species composition). Our results indicate that multiple disturbances have impacts on the structure and functioning of tropical forest, with important implications for their postdisturbance recovery trajectories. These impacts lead to the homogenization of tree communities—a process driven by the increase in generalist species (i.e., a “generalization process”)— but also providing some evidence of strong forest resilience when disturbance is removed from the environment. This understanding is vital for effective forest conservation and strategies for mitigating forest degradation in the face of ongoing global changes.

Materials and Methods

Study Site.

The study site was located at the Tanguro Research Station on Fazenda Tanguro (83,000 ha), a private farm in Mato Grosso state, Brazil (13°04′S, 52°23′N) (SI Appendix, Fig. S1). Located in the transitional zone between the Amazon and Cerrado, the region is characterized by a complex mosaic of seasonal forests and savannas. The study area vegetation is composed of seasonally dry forest, with an average height of 20 ± 1 m [mean ± SE (31)]. The region is characterized by a prolonged dry season (4 o 5 mo) from May to September, with annual precipitation of 1,770 mm (50). The region’s soils are predominantly yellow-red dystrophic latosols, characterized by low fertility and high sand content (51).

Half of the area of Fazenda Tanguro is primary forest, while the remaining area has been cleared for agriculture, beginning with cattle ranching in the 1970s followed by crop production in the 2000s (mainly soybeans, corn, and more recently cotton). The experimental fire plots are adjacent to a field that was initially planted with exotic pasture grasses, which now grows in rotation with crops. The experimental site is composed of three adjacent 50-ha plots: one plot burned annually (B1yr—except 2008), another triennially (B3yr—2004, 2007, and 2010), and an unburned (Control). The plots were burned from 2004 to 2010 at the peak of the dry season (August and September) with the application of fuel for initial ignition. Two regional droughts (low rainfall and relative humidity) co-occurred with the prescribed fires of 2007 and 2010, which promoted higher intensity and high-impact fires. In addition, in 2012 (e.g., ref. 25) and 2019, windstorms caused significant damage to forest structure, particularly along forest edges. Several studies have presented the results of fire behavior (9, 27, 31) and fire-related impacts on forest structure and functioning (30, 52).

Vegetation Sampling.

We performed a forest inventory in all the three treatments before (2004) the prescribed fire period and nine times during and after the initial prescribed fires (SI Appendix, Fig. S1). In 2004, before the initial burn, we mapped, tagged, and measured the diameter at breast height (dbh: cm) and height (m) of trees, palms, and lianas (here all called “trees”; see SI Appendix, Table S2) according to different sampling strata (31). We measured and mapped all trees with a dbh ≥ 40 cm present in each 50-ha treatment plot. Then, we constructed six transects parallel to the agricultural field at 0, 10, 30, 100, 250, 500, and 750 m from the edge to sample smaller trees (SI Appendix, Fig. S1). On the southern side of each transect, we established nested subplots at 2.5, 5, and 10 m (SI Appendix, Fig. S1). We constructed subplots starting at 0.5 m from the transect to prevent the inclusion of trees impacted by the transects themselves. On the northern side of the transects from the edge (except transect A; SI Appendix, Fig. S1), the second (5 m), and third (10 m) subplots were also constructed. Within the subplots we sampled all the trees according to the inclusion criteria adopted. For 2 m (2 × 500 m; 0.7 ha sampled per treatment plot), we sampled trees 5 to 10 cm dbh. In the subplots 10 m (10 × 500 m; 3.25 ha sampled per treatment plot), we sampled trees 10 to 19.9 cm dbh. Finally, in the subplots 20 m (20 × 500 m; 6.0 ha sampled per treatment plot), we sampled trees 20 to 39.9 cm dbh. Unlike the larger diameter classes, trees 5 to 10 cm dbh were only included in the inventories starting in 2006; the other size classes were first surveyed in 2004. In subsequent years, we repeated the inventories within each sampling stratum (2008, 2010, 2012, 2014, 2016, 2018, 2022, and 2024). All mapped living individuals and recruits that reached the inclusion limit within each sampling stratum were measured during resampling.

Tree species were also classified according to their successional status using functional criteria (e.g., shade intolerant or tolerant; SI Appendix, Table S2). We classify shade-intolerant species as pioneer trees species or “fast-growing species” that establish in deforested or disturbed areas (e.g., canopy gaps), persisting only under initial conditions of high light availability. Shade-tolerant or “late-successional” tree species, on the other hand, preferentially grow under lower light conditions, exhibit slow growth, and are strong competitors (53). To classify the successional group for each species, we consulted specialized literature and experts (SI Appendix, Table S1).

Plant Functional Traits.

We collected all other traits at the peak of the rainy season in the region (January to March 2019), which represents the peak of leaf expansion and maturity. We sampled 2 to 5 individuals per species within each treatment. The trait data related to fire resistance, resource use, and water stress were obtained between February and March 2019. At the same time, the hydraulic traits were obtained between August 17 and September 10, 2020 (water potential at midday and predawn at the peak dry season) and March 2021 (embolism vulnerability curves and HSM).

Fire resistance.

Following previously published methods (15), we sampled two functional traits related to the fire resistance of 91 species (SI Appendix, Table S2): wood density (WD) and bark thickness (BT). According to the protocol, we removed the inner and outer bark with a chisel and measured absolute bark thickness using a digital caliper. The individuals for which bark thickness was measured had a mean diameter of 19.2 cm (±14.5 cm), with values ranging from 5 to 110 cm. Consequently, our measurements do not reflect ontogenetic increases in bark thickness associated with individual tree growth through time. To calculate the wood density, we used the dimensional method, which consists of obtaining the cylindrical volume of the branches using calipers, and then dividing it by the sample’s dry weight. As these data were collected in 2019, we calculated CWM of these traits by relative abundance for forest assemblages over the period 2004-2024, assuming that the measurements were representative of climate conditions throughout the entire study period.

Resource use and water stress.

According to the protocol (15), we sampled two functional traits related to water stress of ~91 species (SI Appendix, Table S2): SLA and maximum plant height (H_max). To obtain SLA, we scanned the leaves and used the EBImage package in R to quantify the leaf area of each image. We recorded the heights of all individuals sampled during the inventories conducted over the years. To obtain the maximum height of each species, we plotted their height distribution across all treatments and recorded the value at the 95% position. We adopted this methodology to avoid possible outliers in plant height values. Similarly, we calculated the weighted averages of these traits by relative abundance, assuming the measured values applied to the entire study period.

Hydraulic traits.

To evaluate how multiple disturbances influenced forest vulnerability to drought, we measured various hydraulic traits of tree species in the experimental area between 2020 and 2021. We then calculated the weighted averages of these traits by relative abundance for forest assemblages in 2004 and 2024, assuming the measurements conducted between 2019 and 2020 were representative of the conditions prevailing throughout the entire study period. Specifically, we measured the following hydraulic traits: leaf water potential (Ψ_midday and Ψ_predawn), embolism vulnerability (P₅₀), and HSM. The leaf water potential (Ψ) indicates the water status of the plant, measuring the tension at which the xylem transports water, which reflects a ratio between atmospheric evaporative demand and water availability in the soil (54, 55). Leaf water potential measurements taken in the early morning (Ψ_predawn) indicate the plants’ ability to access water via deep roots in the soil (56). The minimum leaf water potential (Ψ_midday) represents the level of water stress that each plant can endure under adverse conditions (11, 57). We used P₅₀ as an indicator of xylem resistance to hydraulic failure—the xylem water potential at which 50% of hydraulic conductance is lost. We also used HSM as an indicator of species vulnerability to drought—defined as the difference between the minimal seasonal water potential and the threshold at which 50% loss of hydraulic conductivity occurs, representing a measure of drought resistance and exposure in plants (56).

For measuring the hydraulic traits, we collected one branch about 1 m long from three to five individuals per species. We use the pneumatic method to calculate the vulnerability to embolism (58). This method measures the relationship between the percentage loss of xylem conductivity [PLC, estimated from the percentage of air discharged (PAD)] and the xylem water potential (Ψ_x), which were collected simultaneously using a pressure chamber (PMS 1000; PMS Instruments Co.). To quantify the embolism curve, the branches were bench dried repeatedly to obtain different water potentials and their respective air discharge (59). To obtain the xylem water potential (Ψ_x), we estimated the leaf water potential (Ψ_l) after keeping the leaves in a black plastic bag for about 20 min to rebalance the water potential of the leaves with the xylem. The percentage loss of conductance was derived from the percentage air discharge curves, and the P₅₀ value was obtained by fitting a sigmoidal curve that linked the percentage loss of conductance to xylem water potential. We fitted curves for each species by aggregating all individuals of a given species. For both the P₅₀ and HSM values, we used the Bayesian approach to better fit the model curve. We estimated species-level P₅₀ and HSM using a Bayesian hierarchical model implemented in the “rethinking” package in R (60). Sigmoidal vulnerability curves were fitted to individual-level PLC data, allowing both HSM and P50 to vary by individual, with species-level estimates derived from the posterior distributions. This approach accounts for variation among individuals and propagates uncertainty throughout the estimation processes. Using the Bayesian approach, we can allow the inference of parameters from observations via the likelihood (61).

To calculate the HSM, we collected branches with mature leaves at the peak of the dry season (from September 10 to October 10, 2020) between midday and 2:00 pm and measured the minimum leaf water potential (min Ψ_leaf) using a pressure chamber (PMS 1000; PMS Instruments Co.). For each species, we calculated the HSM as the difference between the minimum water potential and the P₅₀ values.

Species Occurrence and Climatic Niches.

For the 122 tree species sampled in our experiment (SI Appendix, Table S2), we compiled occurrence records (geographical coordinates) in the Amazon and Cerrado region from three online databases: i) SpeciesLink (62), ii) Brazilian Biodiversity Information Facility Repository—SiBBr [https://www.sibbr.gov.br/ (63)], and iii) the Global Biodiversity Information Facility—GBIF [https://www.gbif.org/ (64)]. After data processing, removal of duplicates, and cleaning, we obtained a dataset of 44,750 records. These data were used to test the relationship between changes in species composition and their precipitation niches, based on species habitat preference index (SI Appendix, Fig. S28), as well as the species climatic niches.

To determine species habitat preference, we followed the methodology of Flake et al. (17). Using information contained in the “locality” and “notes” fields, we searched for the terms “campo,” “cerrado,” “savanna,” and “grassland” (including variations) to identify savanna occurrences, and for “mata,” “floresta,” “cerradão,” “bosque,” and “forest” to identify forest occurrences. Based on these records, we classified species habitat preference as proportion of occurrences recorded in forest habitats, where 0 indicates savanna specialists (occurring predominantly in savanna habitats) and 1 indicates forest specialists (occurring predominantly in forest habitats). Species classified as generalists tend to occur primarily in forest habitats but are also present in savannas.

To describe the climatic niche of each species, we used the MCWD. We derived average monthly MCWD from monthly CHIRPS precipitation data (65) from 1981 to 2020 for all species occurrence points in the Pan-Amazon and Cerrado. The MCWD represents the maximum accumulated monthly water deficit (mm) during the driest period, providing a measure of species drought tolerance. We extracted the MCWD values from climate databases that reflect the hydroclimatic conditions of regions where species occur, with values close to zero indicating species adapted to humid regions and more negative values indicating greater tolerance of aridity (16, 19). We tested the relationship between species composition changes and their precipitation niche and species habitat preference by weighting their relative abundances based on our forest inventories over time using the CWM approach.

Statistical Analysis.

We used generalized linear mixed (66) and additive (67) models to test for differences among treatments over time. These models included the response variable of interest (e.g., plant functional traits) and predictors such as treatment and year. We used the functional trait values of each species and weighted them by relative abundance using the CWM approach. Previous studies at this site (e.g., refs. 27, 28, and 30) showed differences in tree mortality rates and canopy cover between the forest interior and edges of burned plots. Consequently, we included a categorical variable representing forest edge (0 to 200 m into the forest) and interior (>200 m inside the forest) in most of our models.

To estimate CI for CWM trait values, species habitat preference, and climatic niche for each plot (Control, B1yr, and B3yr) and environment (edge and interior), we used a parametric bootstrapping approach with 1,000 iterations. Trait samples for each species were drawn from a normal distribution defined by the observed species-specific mean and SD. In each iteration, CWMs were calculated as weighted averages of the sample trait values, using species relative abundance as weights. 95% for CWMs were derived using the percentile method. The analysis was implemented in R, with functions customized to calculate CWMs.

We compared species richness across treatments plots and environment (edge and interior) over the years applying rarefaction curves, standardizing the sampling effort by the number of individuals in each treatment (26) using the iNEXT function from the iNEXT package in R (68, 69). To identify potential differences across the distinct phases of the experiment (i.e., Prefire, Fire, and Recovery), we constructed rarefaction curves for each treatment plot. To further compare changes in tree species composition (for all DBH classes) between the treatments plots, we employed NMDS using the Bray–Curtis dissimilarity index (70) and tested the significance of these compositional changes with an ANOVA. Since smaller trees (>5 cm DBH) were sampled only from 2006 onward, all analyses were restricted to trees >10 cm DBH, except for the NMDS, which included all DBH classes.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

Inventory data have been deposited in Zenodo (71–73).