Windthrow characteristics and their regional association with rainfall, soil, and surface elevation in the Amazon

The study area is the Amazon rainforest (5.65 × 10⁶ km²) [18]. We used 392 Landsat 8 images (each image covering ∼3.4 × 10⁴ km², and 30 m × 30 m of spatial resolution) with low cloud cover (<20%) from 2018 and 2019 (see supplementary figure S1). Windthrows were identified by their spectral characteristics [17] and distinctive fan-shape diverging from a central area [35]. To identify windthrows, spectral mixture analysis (SMA) [43] was applied to bands 2–7 in Landsat 8 images. SMA quantifies the per pixel fraction of endmembers which sums to match the full pixel spectrum of the image [43]. Image-derived endmembers of photosynthetic vegetation (PV), non-photosynthetic vegetation (NPV), and shade were used. One set of endmembers was used to identify windthrows across the whole Amazon (figure S2). The fractions of PV and NPV were then normalized without shade as PV/(PV + NPV) and NPV/(PV + NPV) [44]. The normalized NPV and PV images were used to identify new and old windthrows, respectively. Each identified windthrow was verified using historical Landsat images dating back to 1984. Windthrows were visually verified for their fan-shape (diverging from a central area with small pixels scattered at the tail). For this task we used a band combination RGB as shortwave infrared (1.57–1.65 μm), near infrared (0.85–0.88 μm) and red (0.64–0.67 μm). In this band composition, old windthrows look bright green (due to the presence of pioneers that reflect more near infrared [45]) and new windthrows (<1 year old) look red (due to high shortwave infrared reflectance from exposed bare wood) (figure S3). The 2018–2019 Landsat images used encompass new windthrows as well as old windthrows in the Amazon for a period of ∼30 years since the Landsat reflectance recover to pre-disturbance conditions in about 40 years in the central Amazon [45, 46] and about 20 years in the western Amazon [47]. We only considered windthrows that were further than 5 km from human settled areas. In total we identified 1116 windthrows.

To determine windthrow area, the main gap of each windthrow was manually delineated by a triangle with one vertex in the beginning of the fan-shaped windthrow and the other vertices at the extreme ends of the fan (figure S3). The area of the windthrow was determined by the area of its respective triangle. The selection of Landsat 8 images (Landsat 8 Collection 2 Tier 1 calibrated top-of-atmosphere, TOA, reflectance) with less than 20% of cloud cover, the SMA analysis, the delineation of the windthrows area, and the calculation of their area with their respective centroid were done in the Google Earth Engine (GEE) platform [48] that provides tools and algorithms for these tasks. Only windthrows larger than 5 ha were used in our analysis. The direction of the windthrow was determined from the vertex (tail) representing the beginning of the fan-shaped windthrow to the middle of the side connecting the other vertices (figure S3). We used the geographic convention of true north as 0° with the angle increasing clockwise.

2.2.1. Rainfall data

2.2.2. Soil classification

2.2.3. Soil properties

2.2.4. Elevation

2.2.5. Landform classification

2.2.6. Temperature

2.2.7. Canopy height

2.2.8. Wind speed

2.3.1. Density of windthrow occurrence and size

The environmental data were resampled to 0.1°, to match the windthrow density resolution. The resampling used the cubic spline resampling method on all datasets except for the Soil Classification data which was resampled using the nearest neighbor method, a common practice for categorical data [70]. The resampling was done in QGIS. The spline resampling method produces reliable resampling results for climate [72] and topographic [73] data. The environmental data used here are the best data currently available for the Amazon and properly represent the climatic, edaphic, topographic, and forest structural characteristics observed in the Amazon [74]. The availability and continuous improvement of these data are key to understanding the dynamics of this ecosystem.

2.4.1. Spatial autocorrelation, regionalization, and spatial regression

The highest windthrow density (figure 1(A)) was observed in the northwestern (Peru and Colombia) and central (Brazil) Amazon, and regions with lower occurrence were observed in the northern (Venezuela, Guyana, Suriname and French Guiana), eastern (Brazil) and southern (Peru, Bolivia and Brazil) Amazon. The largest windthrows were observed in the northwestern Amazon followed by the central Amazon (figure 1(B)). We found that most windthrows (81% of cases) have a westward component (figure 1(C)). A map with the direction of each windthrow is shown in figure S5.

Zoom In Zoom Out Reset image size

Using regionalization analysis, we found the highest CH score (table S2) was using windthrow density, Elevation, SOC stock from 0–5 cm depth, and MAR (figure 2). The Moran's I value was >0.7 (p = 0.001) for these variables suggesting clear spatial structures with non-random distributions.

Zoom In Zoom Out Reset image size

The Queen weighting method yielded the best regionalization results (table S2). The four Regions obtained using this method are shown in figure 3 (results using the KNN method are in figure S6). The distribution profile of each variable for each Region is shown in figure S7. The number of windthrows contained in Regions 1, 2, 3, and 4 were 390, 291, 297, and 44, respectively. Using 5 regions we obtained CH scores of 256 and 277 using the Queen weighting and KNN methods, respectively.

Zoom In Zoom Out Reset image size

Using the results from our regionalization analysis, we determined an association between windthrow density, the predictant, and predictors (Elevation, SOC stock from 0–5 cm, and MAR). No multicollinearity was observed in the models (table S3). The F-statistic showed that all models were significant and have predictive capabilities (probability < 0.05), except for Region 4 (probability = 0.5) (figure 4). The errors from the models with respect to observations were normally distributed around zero (figure S8), indicating no critical under- and over-estimates. The results showed that windthrows are positively correlated with MAR (except in Region 2), but vary inversely with Elevation (except in Region 4) and SOC. Overall, MAR and Elevation contribute the most (largest magnitude) in the prediction of windthrows. Across the whole Amazon the predictors explained 11% of windthrows density (figure 4(E)).

Zoom In Zoom Out Reset image size

Windthrows were common from the rainy northwest to the dry southeast Amazon. The high occurrence of windthrows in the central Amazon is consistent with a recent study of small forest gaps in this region [40] and was not identified in previous studies [35, 36] that used a threshold windthrow size of 30 ha which might have underestimated the occurrence of windthrows. Only one study has focused on the spatial-temporal variability of windthrows in the central Amazon for the period 1998–2010 [84]. That study did not find an increase of windthrows, suggesting that if an increase has occurred it has happened after 2010. Nevertheless, an increase in windthrows would be consistent with the observed increase in tree mortality found in inventory plots across the Amazon [5] and could provide insights into the mechanisms underlying this trend. An increase in the frequency of windthrows in the central Amazon could promote changes in the regional floristic composition as shown in Negron-Juarez et al [18]. Our study included only windthrows >5 ha but windthrows cover a large spectrum of disturbances including single downed trees. Integration of field and remote sensing data (e.g. LiDAR, Landsat, etc), is needed in order to determine the fraction of tree mortality associated with winds in the Amazon. Such integration will provide a better understanding of the Amazon. For instance, by including wind induced tree mortality, we found that the forest dynamics associated with these events properly represent the values of biomass in the central Amazon [16]. Our results also showed a low number of MANERR in the northeastern Amazon (1.5° N, 53° W) (figure S4), an area where recent studies have found trees taller than 80 m [85].

Our study showed that windthrows in the Amazon have a preferential east to west direction (figure 1(C)). This direction coincides with the dominant direction of MCSs in the Amazon [27]. On the other hand, windthrows with an easterly direction are less frequent and can be related to the lower frequency of easterly MCSs [84]. These results suggest that MCSs are key drivers of windthrows and complement previous studies that found squall lines (i.e. aligned bands of MCSs) [86] to be the drivers of windthrows [17, 29, 84], despite the fact that squall lines are not common in the northwestern Amazon [86], a region where we found a large amount of windthrows.

Regions 1 and 3 were characterized by elevations < 300 meters above sea level and a lower number of extreme rainfall events than the western Amazon (figure 2(B)). These Regions appear to have the most severe storms across the Amazon as suggested by the high values of convective available potential energy [87], and more frequent lightning [88]. Observational studies have shown that during the wet season, rainfall occurs independent of elevation; however, differences on the order of 100 m in surface elevation could provide the forcing for cloud formation during the dry season [89]. Regions 2 and 4 are located in the northwest part of the Amazon, an area characterized by higher elevations, due to its proximity to the Andes. These two Regions had the largest values of MAR (and MANERR) because higher elevation favors higher rainfall. This is due to the subsidence of thermal heating which allows an increase in convection that organizes into MCSs [90]. However, the closer a region is to the Andes, the lower the rainfall, as rainfall decreases with elevation [91]. The patterns we report here require further investigation spanning temporal scales since trends in extreme rainfall have been observed across the Amazon due to deforestation and climate variability [92].

Windthrow correlated positively with MAR. In general, rainfall can promote saturated soils that in turn limit the anchoring of trees making them more sensitive to winds [18]. Regions 2 and 4 have a high frequency of extreme rainfall events (figure S4), and a high occurrence of windthrows. Coincidently, in these Regions studies have shown a low tree residence time (high tree mortality) [22]. On the other hand Regions 1 and 3 appear to have the most severe storms [88] which might help to explain the high occurrence of windthrows in these Regions. In Region 2, maximum MAR values were observed in an area with low elevation in the border between Colombia, Brazil and Venezuela where the relief shape and concavity of the Andes promote air convergence and high rainfall [93, 94]. Associated with this high rainfall is cloud cover (figure S1) that might have influenced the low number of identified windthrows along the border between these countries.

Windthrows vary inversely with elevation. A decrease in tree mortality with elevation was found in the tropical forest of Borneo [95]. A decrease in tree mortality with elevation was also found in inventory plots in other tropical forests [42, 96]. Furthermore, observational studies have shown that wood density increases with elevation [95, 97–99], and therefore these trees are less vulnerable to winds [99, 100]. Similarly, trees in plateaus have deeper root systems [101, 102] resulting in stronger tree anchoring. These studies suggest that, apart from tree-soil interactions and differential exposure to wind gusts, variations in tree mortality across topography are likely to reflect the vertical heterogeneity of the forest canopy, as well as interactions among trees, lianas, and epiphytes, which can influence the mechanical stability and anchoring efficiency. In Region 4 windthrows vary positively with elevation but this effect is very small and was not significant. Our results corroborate the important and complex role of elevation on the occurrence of windthrows.

Windthrows vary inversely with SOC. The occurrence of windthrows implies the presence of dead trees and therefore SOC [59, 103] and nutrient availability [104] dependencies. Studies have shown that large amounts of rainfall promote accumulation of SOC [105, 106], explaining the large amounts of SOC in the northwestern Amazon (figure 2(C)). A decrease in SOC is observed from Regions 2 and 4, in contrast to the increases in windthrows in those regions (figure 1(B)). Similarly, an increase in SOC is observed from east to west in Regions 1 and 3 (figure 2(C)), in contrast to a high density of windthrows in the eastern-central part of the Amazon (figure 1(A)). These patterns explain the negative association between windthrows and SOC (table 1). Previous studies suggest that SOC, nitrogen (N) and phosphorous (P), are related [39, 107], suggesting that the effect of N and P on windthrows might be similar to that of SOC.

Windthrows have the potential to change forest composition in the Amazon [18]. Forest composition influences the climate feedback from cloud formation and rain to carbon uptake [108–111]. Since rainfall has a direct effect on the occurrence of windthrows, we suggest that Earth system models should include the effects of windthrows to reduce uncertainties on the projected functionality of the Amazon in the future.

Intrinsic to windthrows is the fact that forested surfaces facing the direction of the storms have the most windthrown trees as shown in our previous studies [17, 18, 112, 113]. Westward MCSs will have a larger impact on surfaces with an eastern aspect, and eastward MCSs will affect largely western-facing surfaces. Thus, it has been suggested that aspect (the direction that a terrain surface faces) is not an explanatory variable, but rather a factor [114]. Aspect in the Amazon is a local characteristic that can produce misleading results after upscaling since several aspect values will be averaged together, muting their effect [115]. We estimate that about 10%–15% of all windthrows identified did not have a clear fan shape, and were not included in the total number of windthrows identified (1116). A more comprehensive regression or machine learning analysis should explore the effect of endogenous variables as well as different regimes in each of the selected regions. Regimes analysis implies that in each Region the regression model changes due to the spatial heterogeneity of the data [82]. We found that windthrows in the Amazon have regional characteristics of occurrence and size, and further analyses are needed to understand the causes of these characteristics.