Floodplains as an Achilles’ heel of Amazonian forest resilience

Climate change may alter the distribution of biomes in tropical regions with implications for biodiversity and ecosystem services. Here we reveal that if the Amazon region becomes drier as predicted, forests may collapse first on seasonally inundated areas due to their vulnerability to wildfires. The widespread distribution of floodplain forests at the western and central regions implies that fire-prone savannas may expand deep into this massive forest biome, threatening the resilience of the entire system. Our findings suggest the need for a strategic fire management plan to strengthen Amazonian forest resilience in the face of climate change.

Abstract

The massive forests of central Amazonia are often considered relatively resilient against climatic variation, but this view is challenged by the wildfires invoked by recent droughts. The impact of such fires that spread from pervasive sources of ignition may reveal where forests are less likely to persist in a drier future. Here we combine field observations with remotely sensed information for the whole Amazon to show that the annually inundated lowland forests that run through the heart of the system may be trapped relatively easily into a fire-dominated savanna state. This lower forest resilience on floodplains is suggested by patterns of tree cover distribution across the basin, and supported by our field and remote sensing studies showing that floodplain fires have a stronger and longer-lasting impact on forest structure as well as soil fertility. Although floodplains cover only 14% of the Amazon basin, their fires can have substantial cascading effects because forests and peatlands may release large amounts of carbon, and wildfires can spread to adjacent uplands. Floodplains are thus an Achilles’ heel of the Amazon system when it comes to the risk of large-scale climate-driven transitions.

The sensitivity of the Amazon rainforest to climate change is a central issue in global change research (1, 2). In particular, there are concerns that a drier climate may promote a shift from forest to savanna (3, 4). All studies so far suggest that this is most likely at the southern and eastern peripheral regions where precipitation is relatively low and seasonal (2–5) and the risk of wildfires is higher (4, 6–11). Although rainfall is a dominant factor explaining forest resilience, other environmental factors obviously play a role (12–14). Arguably, the most striking variation in the nature of forests in the wet Amazonian system is related to seasonal inundations. Approximately one-seventh of the Amazon is inundated a substantial part of the year (15), causing these ecosystems to differ in many ways from the dominant upland terra firme forests (SI Brief Ecology of Floodplain and Upland Ecosystems). Here we ask whether these differences related to seasonal inundation affect forest resilience and the risk of shifting into a fire-dominated savanna state. We used two approaches to contrast the resilience of floodplain and upland forests across the Amazon (Fig. 1 A and B). First, we estimated the long-term relative resilience of forest and savanna in both ecosystems from the density distributions of tree cover (3, 16) using MODIS vegetation continuous field (VCF) data at 250 m resolution (SI Methods). Second, we studied postfire recovery of both forest types using field and remote sensing data (Fig. S1). Using annual MODIS VCF data, we measured the recovery of over 250 sites that burned during the severe droughts of 1997 and 2005 (Fig. S1; Table S1; and Dataset S1). Using field data on tree basal area and soil variables from multiple burned forests in the central Amazon region, we validated the basin-wide analyses of postfire recovery (Fig. S1 and SI Methods).

Fig. 1.

Distribution of tree cover across the Amazon basin. (A) Floodplains and (B) uplands separated by the wetlands’ mask (15). Deforested areas were excluded (*SI Methods*). Density distributions of tree cover for (C) floodplains and (D) uplands. A cutoff at 60% tree cover (dashed lines) separates forest from savanna (3), and percentage values are the proportion in each state. Relation between mean annual rainfall and tree cover for (E) floodplains and (F) uplands. Markham seasonality index values >36 (black circles) and <36 (gray circles) are shown. Red lines are fitted locally weighted scatterplot smoothing (LOWESS). Circles are 15,000 data points randomly sampled for each case.

Fig. S1.

Maps showing sites selected to study wildfires with satellite and field data. (*Left*) Tree cover for South America. (*Right*) Floodplains and mean annual rainfall. Black squares show Landsat scenes chosen to study the impact of wildfires on floodplains and uplands. Locations of field sites used for the fire study are indicated by red circles for uplands and blue circles for floodplains.

Table S1.

Number of sites selected in each Landsat scene

Region	Landsat scene	Ecosystem	Burned in 1997	Burned in 2005	Unburned
South	230/69	Upland	3	9	6
		Floodplain	6	22	6
Southeast	225/68	Upland	10	16	6
		Floodplain	5	5	6
Southwest	001/68	Upland	21	48	6
		Floodplain	4	37	6
North	232/59	Upland	39	0	6
		Floodplain	21	0	6
Central	231/62	Upland	12	0	6
Central-North	233/60	Floodplain	11	0	6
East	226/61	Upland	0	0	6
		Floodplain	0	7	6
West	006/63	Upland	0	0	3
		Floodplain	4	6	6
Amazonia		Upland	85	73	39
		Floodplain	51	77	42

In these sites, we studied tree cover change after the 1997 and 2005 fires (Fig. 2 and SI Methods).

The rationale behind studying the density distributions of tree cover is that under homogeneous environmental conditions (14), density distributions may reveal alternative attractors in the vegetation and their relative resiliencies (3, 16, 17). Comparative studies across the global tropics (3, 4) reveal that tree cover tends to be either high or low, with intermediate (∼50%) cover being remarkably rare. The interpretation is that closed forest and savanna are attractors, and the intermediate state is an intrinsically unstable repellor. In fact, one may reconstruct the well-known potential (ball-in-cup) landscapes directly from the data. The mathematical underpinning is somewhat technical (17), but the idea is intuitively straightforward. Thinking of long time-spans, one can imagine that stochastic events occasionally push the system over the boundary between the basins of attraction of alternative attractors. Sampling such a system long enough, one thus expects it to be sometimes close to one of the attractors (e.g., forest), sometimes close to the alternative attractor (e.g., savanna), and more rarely somewhere in between (16). Importantly, in places where forest is relatively more resilient, the system will spend more time in the state with high tree cover, whereas in places where savanna is relatively more resilient, the system will spend more time in the state with low tree cover. Thus, the ratio of the number of observations in the forested state vs. the number of observations in the savanna state tells us something about the relative resiliencies of the two alternative states. We do not have long time-series, but if information for sufficient sites of comparable conditions is available, one may interpret snapshots of the states on all sites in the same way as one would interpret samples from a single, long time-series. Such a space-for-time substitution has been used before to show how resilience of tropical forest and savanna varies with mean annual precipitation (3). We now use the same approach to infer how resilience of forest and savanna varies between floodplains and uplands of the Amazon.

SI Brief Ecology of Floodplain and Upland Ecosystems

One-seventh of the Amazon basin is formed by seasonally flooded ecosystems (Fig. 1) (15). The larger area is covered by nonfloodable terra firme ecosystems (86%), which we refer to as uplands. Floodplains differ from uplands in many ways, because life has to adapt to an annual flood pulse, altering between the terrestrial and aquatic phases (25, 46). Due to the predictability of such disturbance, life can flourish with special adaptations. Of the 227 hyperdominant Amazonian trees, one-fourth is specialized to seasonally inundated habitats (47), implying that floodplain trees contribute with broad-scale ecosystem functions. Despite inundation, adjacent floodplains and uplands share one-third of their tree species (48), revealing connectivity between their floras.

Floodplain forests across the Amazon have an average aboveground live biomass of 160 tons per ha (±100 of SD) (36). The variation is likely associated with distinct habitats and water types (46). Rivers that carry white water originate in the Andes, from where they transport huge loads of sediment and nutrients; their floodplains are highly productive ecosystems, and represent half of Amazonian floodplains. In contrast, black and clear water rivers are poor in sediments and nutrients, and have floodplains with low fertility. The acid and nutrient-poor black water that flows in countless streams and tributaries born in the forest has a major influence on half of Amazonian floodplains (46). Amazonian floodplains also include peatlands (26, 27), swamps, palm forests, white-sand forests, and extensive islands of savanna (46).

Upland forests have an average aboveground live biomass of 250 tons per ha (36), which varies across habitats (±100 of SD). Basin-wide environmental gradients explain most of the variation in forest structure and dynamics (49). The western Amazon is the region that receives more rain throughout the year, and has more fertile soils due to a major geological process known as the Andean uplift. Around 7 Mya, the Amazon River was formed and began to supply nutrients from the Andes to the western part of the Amazon basin (12), while massive wetlands started retreating, allowing the expansion of new upland habitats. On the other extreme, upland forests of the Guiana Shield region have the poorest soils of the basin. Paleoecological evidence suggests that most of the Amazon remained forested during the last 8,000 y, and that only in ecotonal areas forest and savanna biomes expanded (18).

Results and Discussion

In Amazonian uplands, the density distribution of tree cover has a single mode around 84%, reflecting dense forest, with sparse tree cover being rare (Fig. 1D and Fig. S2B). In contrast, floodplains have an additional mode around 34% tree cover, reflecting the presence of a savanna state (Fig. 1C and Fig. S2A). As explained previously, assuming our density distributions (Fig. 1 C and D) reflect long-term dynamics of a stochastically perturbed system, the ratio of the number of observations falling in the two modes reflects their relative resiliencies (3, 16). Forest/savanna ratios we found suggest that forests are much less resilient in floodplains than in uplands (ratios of 66/34 vs. 93/7; Fig. 1 C and D and Table S2). The difference is especially pronounced in parts of the Amazon where rainfall is relatively lower, more seasonal, and interannually variable (Fig. 1 E and F and Fig. S3). From the relation between tree cover and rainfall, we computed potential landscapes (Fig. S4 and SI Methods), which revealed that a savanna basin of attraction appears around 1,500 mm of mean annual rainfall in floodplains. In uplands, a hint of the savanna basin of attraction only becomes apparent around 1,000 mm of annual rainfall.

Fig. S2.

Probability density distributions of tree cover across the Amazon. (A) Floodplains. (B) Uplands. Percentage values indicate tree cover at the peaks of each mode. The mode that peaks at 67% tree cover in uplands is due to overfitting of the long tail, and likely represents secondary forests. For more information see *SI Methods*.

Table S2.

Proportion of floodplains and uplands covered by forest and savanna (tree cover cutoffs at 60%)

Ecosystem	Forest, %	Savanna, %	Area, km²	Forest area, km²	Savanna area, km²
Floodplain	66	34	816,200	538,692	277,508
Upland	93	7	5,013,800	4,662,834	350,966
All Amazon	89	11	5,830,000	5,201,526	628,474

Areas calculated with wetlands’ mask (15), excluding deforested areas (see SI Methods for information).

Fig. S3.

Relation between climatic variables and tree cover. (*Left*) Floodplains. (*Right*) Uplands. Black lines are fitted LOWESS smoothing. Circles are 15,000 data points randomly sampled for each case as in Fig. 1 E and F.

Fig. S4.

Computed landscapes of probability densities for the Amazon based on potential analysis (17). (A) Floodplains. (B) Uplands. Blue areas approximately represent the basins of attraction; black dots represent the local minima; and white dots the local maxima, as the stable and unstable states, respectively. More information in *SI Methods*.

Thinking of our tree cover density distributions as the long-term balance between shifts from forest to savanna and vice versa, the forest/savanna ratio should approximately reflect the average time that the system spends in each state (16). Long-term tree cover time-series are lacking, making it difficult to check this inference systematically. An alternative is to analyze the response of the system to stochastic perturbations. Wildfires have been the dominant historical mechanism driving shifts from forest to savanna in times when climate was drier and seasonal (4, 18). The megadroughts of 1997 and 2005 invoked a large number of wildfires in Amazonian upland and floodplain forests (6–9), allowing us to compare their resilience. Our remotely sensed information from over 250 burned forest sites reveals that fires had a strong and long-lasting impact on tree cover in the floodplains, whereas in the uplands, effects of fire were typically small and ephemeral (Fig. 2 A and B and Fig. S5). This broad-scale pattern was confirmed by field measurements in burned forests of the central Amazon region, showing a systematically lower recovery rate of tree basal area in the floodplains (Fig. 2 C and D). We also found a marked decrease in soil nutrients and fine clay particles in floodplain forests upon fire that was absent in the uplands (Fig. 3).

Fig. 2.

Compared sensitivity of floodplain and upland forests to fire. Time series of annual tree cover (median ± SE) showing changes after a fire in (A) floodplains and (B) uplands. After fire (time = 0), tree cover median persists below 50% in floodplains, yet recovers in uplands. Field data on tree basal area recovery after the last fire in (C) floodplains and (D) uplands of the central Amazon. For all plots, green shaded area is SE intervals for unburned forests. In D, unburned reference was obtained from ref. 40 (*SI Methods* for details).

Fig. 3.

Changes in soil texture and fertility that initiate with a forest fire. (*Left*) Floodplains. (*Right*) Uplands.

Fig. S5.

Example of study sites burned in 2005 shows how quickly signs of fire disappear. Landsat images from Bolivian rainforests (scene 001/68, pseudocolor combination of bands 5, 4, and 3 for channels red, green, and blue). Forests are shown in green and burned areas in magenta. Dashed lines indicate floodplains (blue) and uplands (yellow).

In summary, the observed difference in density distributions of tree cover indicates that floodplain forests are relatively less resilient and therefore less likely to persist in the long run, which does not imply that upland forests are resilient, but rather that their resilience depends on overall environmental conditions. In fact, our results suggest that at ∼1,000 mm of mean annual rainfall, upland forests reach a tipping point in which they may collapse into a savanna state (Fig. S4). As climatic conditions change, however, such tipping points will likely be reached first by forests on seasonally inundated areas. Our detailed field and remote sensing measurements of the response to fire confirm this pattern and suggest an explanation. Both the slow forest regrowth and the quick loss of soil fertility plausibly make these floodable forests more likely to be trapped by repeated fires in an open vegetation state. Indeed, studies in other forest systems reveal that canopy openness upon the first fire enhances the risk of subsequent fires (19) and the spread of herbaceous vegetation (7, 10, 11). The loss of tree cover to fire allows the intensification of hydrological processes that lead to soil erosion and nutrient leaching (20), creating limiting conditions for forest regrowth that may instead favor savanna-adapted plant species (21). Another reason that makes floodplain forests more vulnerable is their naturally higher flammability compared with upland forests. In addition to having a slightly more open structure (22, 23), upon the annual retreat of the waters, floodplain forests typically have large masses of exposed root mats (24) that burn easily and may spread fire effectively in drier years (6, 7) (Fig. S6). The combustion of this organic material by fire may plausibly facilitate subsequent floods to wash away nutrients and fine sediments, leaving behind relatively poor sandy soils. Overall, forest recovery upon fire in the floodplains may be hindered by a combination of recruitment limitations partly caused by lost soil fertility (Fig. 3) with seasonal inundation that restricts the time in which trees can grow (25).

Fig. S6.

Floodplain forests and fire. (A) Floodplains of the Jari River, tributary of the Purus subbasin; (B) root mat is ubiquitous above the soils of blackwater floodplain forests; (C) fireman trying to stop a fire spreading through the root mat in a young successional floodplain forest at the Anavilhanas National Park, Negro river, during the extreme drought of late 2015; (D) Floodplain forest 3 y after a first fire at the middle Negro river. Photos courtesy of Diogo Lagroteria (A); Aline R. dos Santos (B); Enrique Salazar (Chico Mendes Institute for Biodiversity Conservation/National Park Anavilhanas) (C); and Bernardo M. Flores (D).

Independent of the precise mechanisms slowing down forest recovery, patterns we revealed imply strong evidence that Amazonian forests on floodable terrains have a lower resilience when it comes to the risk of transition into a fire-prone vegetation state. Although these ecosystems cover 14% of the basin, there are two reasons why their vulnerability may have substantial cascading effects. First, floodplains in the western and central Amazon include peatlands that store enormous amounts of carbon (estimates of ∼3.14 Pg C, available only for western peatlands) (26, 27), which could potentially be released to the atmosphere by fire. Second, our results imply that if the climate becomes drier, fire-prone savannas might expand through floodable areas toward the core of the Amazon forest and become sources of fires that may spread to large parts of that region. Indeed, the spread of fires from floodplains to adjacent uplands has been shown in Africa (28) and central Amazon (22) with negative impacts on vegetation structure (22) and biodiversity (29). In conclusion, our results suggest that seasonally inundated forests throughout the Amazon represent an Achilles’ heel when it comes to resilience of this massive system. Considering the projected increase in dryness (30) and expected effects of climatic variability intensification on tropical vegetation (31), it follows that maintaining a safe operating space for the Amazon forest (32) may require special protection not only of the driest parts of the Amazon forest, but also of the floodable heart of the system.

Methods

To contrast the resilience of Amazonian floodplain and upland forests, we first analyzed density distributions of tree cover for the entire basin using MODIS VCF data at 250 m spatial resolution (SI Methods). We also analyzed postfire recovery rates as a measure of resilience for over 250 sites in both forest types using annual tree cover data from MODIS VCF at 250 m (SI Methods). To identify forest fires, we selected eight Landsat scenes with extensive areas of floodplain and upland to allow equal probability of observing forest fires. These scenes are distributed across the Amazon basin and cover most of the annual rainfall gradient (1,500–3,000 mm; Fig. S1). In each scene we identified forest sites that were burned during the droughts of 1997 and 2005 using a systematic visual method (33). This method allows distinguishing fire from other disturbances because signs from fire typically have rounded borders and fade away quickly. It has proven to be effective for floodplain forests (6, 7), and we found it to be even more effective on uplands where signs of fire disappear within 3 y (Fig. S5). In each scene we identified multiple sites to accommodate MODIS pixels (250 m), spreading their locations within the burned and unburned areas, maintaining at least 1 km distance between two pixels and from perennial water bodies (see all sites in Dataset S1).

To ground-truth our basin-wide remote sensing analysis of tree cover recovery after wildfires, we analyzed tree basal area and soil data from field sites at the central Amazon region (Fig. S1). In both floodplain and upland areas, we selected a series of forest sites with different “times since the last fire” to produce chronosequences (space-for-time) and measure recovery rates of tree basal area [>1 cm in diameter at breast height (DBH)]. Using these same sites, we produced chronosequences of “time since the first fire” (when mature forests were burned for the first time) to analyze changes that may have occurred in the superficial soil fertility (0–20 cm depth) while repeated fires maintained the vegetation vulnerable to soil erosion (20).

In this study we considered forest and savanna as alternative vegetation states with contrasting structure and plant composition, maintained by positive feedbacks among plants, fire, and soil (34). Different lines of evidence support the use of tree cover data from MODIS VCF to assess forest and savanna distribution and resilience. VCF has been validated in the field with a relationship of 95% with crown cover (35) and 56% with forest biomass (36); more recently it was shown to have a good correlation with field tree cover across the whole tropics (37). Spatial patterns of VCF correlate well with tree canopy height, distinguishing degraded forests from savannas as well as closed savannas from forests (38). Because closed canopies are known to suppress fire percolation, forests that recover faster also recover their capacity to suppress recurrent fires (39); therefore, VCF also indicates spatial variation in flammability. In addition, VCF provides us massive data on forest tree cover at global scales.

Most data sets used in this study are openly available and have been used in other publications. The location of all wildfires studied from remote sensing can be found in Dataset S1. For access to field data from burned upland forests, contact C.C.J. for the Tefé area and R.C.G.M. for the Manaus area. Field data from burned floodplain forests at Barcelos are available from a previous publication (7).

SI Methods

Data Description.

Wetlands mask.

Uplands and floodplains were distinguished across the Amazon basin using the Amazon wetlands’ mask (15) (daac.ornl.gov/LBA/guides/LC07_SAR_Wetlands_Mask.html). This mask identifies areas that are permanently or seasonally covered by water (following the definition of wetland by the Ramsar Convention) at a spatial resolution of ∼90 m with an accuracy of 93%. Floodplain areas were extracted by comparing the remotely sensed images of Japanese Earth Resources Satellite between the low water season of 1995 and the high water season of 1996 when the wetlands’ mask was produced. The mask does not include the Tocantins watershed and most of the Amazon estuarine islands that are part of the lower Amazon basin. Permanent water bodies were excluded.

Tree cover.

The 250-m resolution MODIS VCF Collection 5 dataset for the year 2001 (41) was used to analyze tree cover distributions of Amazonian uplands and floodplains. For the analysis of tree cover recovery after fire, annual MODIS VCF tree cover data between 2000 and 2010 was obtained for each study site (Table S1 and Dataset S1).

Climate.

We extracted climatic data from the Climate Research Unit (42). This high-resolution monthly data set is an interpolation of observations from meteorological stations at the spatial resolution of 0.5°, based on data for the period 1961–2002. With this data we computed mean annual precipitation, the Markham seasonality index, and the precipitation in the driest quarter that indicate how rainfall is distributed throughout the year, and the coefficient of variation that indicates how rainfall varies between years.

Deforested areas.

We excluded all accumulated areas deforested until 2014 obtained from Project Brazilian Amazonian Forest Monitoring by Satellites, a project for mapping annual deforestation in the Brazilian Amazonia run by National Institute for Space Research (www.dpi.inpe.br/prodesdigital/prodes.php). This project mapped annual deforestation since 1988 with the use of Landsat images to be used by the Brazilian government. We also excluded lower Amazon River floodplains because of their long history of human land use (43), and floodplains under tidal influence from the Atlantic coast.

Fire detection.

We analyzed the recovery of tree cover (MODIS VCF) after wildfires in floodplain and upland forests across the Amazon basin. Although signs of fire in the forest canopy can be ephemeral, lasting only 1 or 2 y (Fig. S5), Landsat imagery allows detecting forest fires at 30-m resolution due to clear changes in spectral patterns (6–9). We selected eight Landsat scenes covering most of the annual rainfall gradient (1,500–3,000 mm⋅y⁻¹; Fig. S1). Scenes were selected to include similar proportions of floodplains and uplands, allowing the comparison of fire impact. For each scene we obtained cloud-free images for the years before and after 1997 and 2005, with the date (same month) as close as possible to control for spectral differences related to seasonality. Both years had strong droughts and wildfires penetrated closed forests in different parts of the Amazon basin (6–9). In each Landsat scene, we selected fire scars using a systematic visual identification method (33) to ensure these sites were burned by wildfires. This way, we avoided areas burned intentionally for land use, for instance, which can be easily identified by their straight borders compared with rounded borders produced by wildfires. In the same scenes we also selected sites that did not burn (at least until 2010) as unburned forest references. Burned and unburned sites were chosen to fit MODIS VCF pixels (250 m), maintaining at least 1 km from the border with unburned forests. In large scars, sites were spaced by at least 1 km. All study sites are labeled in Dataset S1 for easy observation (more details in Table S1).

Analyses of Tree Cover Distribution.

We analyzed the frequency distributions of tree cover, and the relationship between tree cover and rainfall in Amazonian floodplains and uplands. The significant number of modes in the tree cover density distribution for floodplains and uplands was tested with latent class analysis that fits normal frequency distributions to the data, using R package FlexMix (R 3.2.1, FlexMix version 3.2.2) (44). The number of modes is quantified based on the most parsimonious model through an expectation–maximization procedure. Bayesian information criterion, the Akaike information criterion, and the integrated completed likelihood criterion were used for model selection. The percent tree cover data were arcsine square root- transformed to approach normal distributions. We randomly sampled 15,000 points for each ecosystem type.

The result of the analyses reveals two modes for floodplains and two for uplands (Fig. S2). For uplands, the mode that peaks at 67% tree cover overlaps with the major forest mode that peaks at 84% tree cover. We interpret this mode as overfitting of the long tail that may reflect deforested or degraded areas that we were not able to exclude completely. Hence we only consider two modes for floodplains and one mode for uplands (Fig. 1), with cutoffs at 60% (3). The computed potential landscapes of floodplains and uplands (Fig. S4) show basins of attraction for each ecosystem based on potential analysis (17).

We also explored the role of the extensive floodplain savanna in the Bolivian south of the Amazon (Llanos de Moxos), and of the extensive floodplain forest in Peru (Marañon-Ucayali basins) in the bimodality of Amazonian floodplains (Fig. S7). When removing the extensive Bolivian savanna, the floodplain savanna mode reduces. Nonetheless, frequencies of tree cover below 60% still remain higher compared with the frequencies observed in uplands (Fig. 2B). Moreover, these extensive floodplain savannas are also at the drier extreme of the annual rainfall gradient (∼1,500 mm⋅y⁻¹), which may indicate a tipping point for floodplain forest collapse (Fig. 1E).

Fig. S7.

Maps and tree cover frequency distributions for Amazonian floodplains. (A) Large forest in Peru (Marañon-Ucayali) excluded; (B) large savanna in Bolivia (Llanos de Moxos) excluded; (C) large forest and large savanna excluded; (D) all floodplains. For information, see *SI Methods*.

Analyses of Tree Cover After Fire.

Using the MODIS VCF annual data (41) between 2000 and 2010, we studied tree cover changes after fires in floodplain and upland forests. Although the 1997 and 2005 fires were concentrated in different regions (Table S1), the response of tree cover was similar (Fig. S8). Therefore, we analyzed both time series as one time series of up to 13 y after fire (Fig. 2 A and B).

Fig. S8.

Annual tree cover (median ± SE) time series for burned forests. (A) Floodplains. (B) Uplands. Green stripe indicates SE intervals for unburned forests in each ecosystem. Data were obtained for the period between 2000 and 2010 for both burned and unburned sites (more information on Table S1 and *SI Methods*).

Interannual MODIS VCF data have been suggested as not appropriate to synoptically document annual change due to the uncertainties arising from clouds and other factors (45). In our case, however, changes in tree cover after fire largely exceed the interannual data variation (Fig. S8), showing that our analysis based on change detection is clearly robust to the data noise in the product.

Field Study of Forest Recovery.

To assess whether the patterns observed with the remote sensing analyses of tree cover recovery after fire could also be confirmed in the field, we analyzed data on tree basal area from secondary forests recovering after being burned in floodplains and uplands of central Amazonia (Fig. S1). For floodplains, we used 15 forest sites burned by one or two wildfires. Three unburned floodplain forests were used as reference. In all floodplain sites, small trees [1–10 cm in diameter at breast height (DBH)] were measured within 300 m²; medium trees (10–30 cm in DBH) within 0.3 ha; and large trees (>30 cm in DBH) within 0.6 ha. These sites are along the floodplains at the Negro river basin within 50 km of Barcelos town. For uplands, we used burned sites from two basins: the Negro and the Solimões. The first basin is located north of Manaus, includes 17 secondary forests abandoned from agriculture, and pasture with fire management, in the area of the Biological Dynamics of Forest Fragments Project. In these upland sites, all trees >3 cm in DBH were measured within 500 m²; the other is located in the lower Tefé river and includes 33 secondary forests abandoned after several cycles of slash-and-burn agriculture. In these upland sites, small trees (1–5 cm in DBH) were measured within 250 m², and larger trees (>5 cm in DBH) within 500 m². For upland unburned forest reference, we extracted tree basal area data from one published review (40). With data from the three field sites, one in floodplains and two in uplands, we used chronosequences (space for time) to analyze the recovery of tree basal area after the last fire. Patterns observed with satellite were then compared with the ones from the field.

Soil Changes After Fire.

We analyzed soil changes after the first fire to test the hypothesis that after the first forest fire that opens the vegetation, soils may leach and erode from hydrological processes (20). Using the same secondary forest sites described above, we produced three new chronosequences of time following the first fire to analyze changes in soil texture and fertility. Superficial soil was collected in subsamples along a centerline in each plot to form compound samples that represented the plot. Analyses were carried out in the Plant and Soil Laboratory (LTSP) of the National Institute for Amazonian Research. A technician from the LTSP analyzed mineral fractions and available nutrients.

Acknowledgments

We thank G. G. Mazzochini for assistance with R scripts for remote sensing analyses, and J. L. Attayde, M. C. X. Flores, C. R. Fonseca, C. Levis, P. Massoca, A. Staal, and E. Venticinque for helpful comments. Support for this work was provided by National Council for the Improvement of Higher Education, Brazil, and the Sandwich Fellowship Program from Wageningen University (to B.M.F.); National Natural Science Foundation of China Grant 41271197 (to C.X.); and an European Research Council Advanced Grant (to M.S. and E.H.v.N.). This work was partially carried out under the program of the Netherlands Earth System Science Centre.

Supporting Information

Supporting Information (PDF)

Download
1.75 MB

Dataset_S01 (CSV)

Download
16.52 KB

References

1

PM Cox, et al., Amazonian forest dieback under climate-carbon cycle projections for the 21st century. Theor Appl Climatol 78, 137–156 (2004).

Featured Topics

Articles By Topic

Featured Topics

Articles By Topic

Featured Topic

Articles By Topic

Significance

Abstract

Sign up for PNAS alerts.

SI Brief Ecology of Floodplain and Upland Ecosystems

Results and Discussion

Methods

SI Methods

Data Description.

Wetlands mask.

Tree cover.

Climate.

Deforested areas.

Fire detection.

Analyses of Tree Cover Distribution.

Analyses of Tree Cover After Fire.

Field Study of Forest Recovery.

Soil Changes After Fire.

Acknowledgments

Supporting Information

References

Information

Published in

Classifications

Copyright

Submission history

Keywords

Acknowledgments

Notes

Authors

Affiliations

Notes

Competing Interests

Metrics

Citation statements

Altmetrics

Citations

Cited by

View options

PDF format

Get Access

Login options

Recommend to a librarian

Purchase options

Restore content access

Figures

Tables

Other

Share

Share article link

Share on social media

Further reading in this issue

Eight-minute self-regulation intervention raises educational attainment at scale in individualist but not collectivist cultures

Spontaneous emergence of catalytic cycles with colloidal spheres

Subduction zone forearc serpentinites as incubators for deep microbial life

AI-accelerated Nazca survey nearly doubles the number of known figurative geoglyphs and sheds light on their purpose

A framework for quantifying individual and collective common sense

Bodily maps of emotions

Sign up for thePNAS Highlights newsletter