Recovery of Forest Structure Following Large-Scale Windthrows in the Northwestern Amazon

: The dynamics of forest recovery after windthrows (i.e., broken or uprooted trees by wind) are poorly understood in tropical forests. The Northwestern Amazon (NWA) is characterized by a higher occurrence of windthrows, greater rainfall, and higher annual tree mortality rates (~2%) than the Central Amazon (CA). We combined forest inventory data from three sites in the Iquitos region of Peru, with recovery periods spanning 2, 12, and 22 years following windthrow events. Study sites and sampling areas were selected by assessing the windthrow severity using remote sensing. At each site, we recorded all trees with a diameter at breast height (DBH) ≥ 10 cm along transects, capturing the range of windthrow severity from old-growth to highly disturbed (mortality > 60%) forest. Across all damage classes, tree density and basal area recovered to >90% of the old-growth values after 20 years. Aboveground biomass (AGB) in old-growth forest was 380 ( ± 156) Mg ha − 1 . In extremely disturbed areas, AGB was still reduced to 163 ( ± 68) Mg ha − 1 after 2 years and 323 ( ± 139) Mg ha − 1 after 12 years. This recovery rate is ~50% faster than that reported for Central Amazon forests. The faster recovery of forest structure in our study region may be a function of its higher productivity and adaptability to more frequent and severe windthrows. These varying rates of recovery highlight the importance of extreme wind and rainfall on shaping gradients of forest structure in the Amazon, and the different vulnerabilities of these forests to natural disturbances whose severity and frequency are being altered by climate change.


Introduction
Windthrows are associated with strong downdrafts produced during severe convective storms [1][2][3][4]. In the Amazon, windthrows are a frequent natural disturbance that influences regional tree mortality and has the potential to regulate regional biomass/carbon stocks and balance [5][6][7][8][9][10]. The geographic distribution of windthrows generally follows the pattern of rainfall across the Amazon Basin, though there are large variations in the size The study area is located within~50 km of Iquitos, Peru, in a region covered with old-growth lowland forest (no flooded areas and less than 500 m a.s.l.) ( Figure 1). Iquitos is characterized by its lack of dry season (no consecutive months with rainfall ≤ 100 mm) [29]. The mean annual rainfall is 3000 mm, and mean annual temperature is 25.9 • C. Precipitation is lowest from June to August with a mean and standard deviation of 183 mm and ±10 mm month −1 , respectively [12]. The study sites are managed by the National University of the Peruvian Amazon (UNAP). Because we want to study the timing of recovery after windthrow disturbances of differing severity, we identified a chronosequence of study sites that were affected by windthrow disturbances at different times in the past. The oldgrowth forests at the study sites have similar climatic, edaphic and floristic compositions. None of our study sites have experienced any major, recent or direct human impacts such as logging.
Windthrow events were identified via Landsat imagery using changes in nonphotosynthetic vegetation (NPV) between sequential years [3,30]. This revealed that disturbances at the three study sites (Nauta, Napo and Oroza), occurred in a chronosequence of 2, 12 and 22 years, respectively, prior to transect installation. Oroza (named after the closest town) was windthrown in 1988, has a disturbed area of 662 ha and was surveyed in 2010 (i.e., 22 years after windthrow). Napo experienced a windthrow event in 1998, has a disturbed area of 188 ha, and was also surveyed in 2010 (12 years after windthrow). The windthrow in Nauta occurred in 2009. This site has a disturbed area of 189 ha and was surveyed in 2011 (2 years after windthrow).
Given that a chronosequence study substitutes space for time [8,31], our study sites are not strictly replicates, and there are differences in forest structure in the undisturbed parts of the transects. We acknowledge that additional temporal measurements would be needed to assess the degree to which this influences our results, and we have, in most Given that a chronosequence study substitutes space for time [8,31], our study sites are not strictly replicates, and there are differences in forest structure in the undisturbed parts of the transects. We acknowledge that additional temporal measurements would be needed to assess the degree to which this influences our results, and we have, in most cases, normalized our data to the respective undisturbed forest when making comparisons across sites.

Assessing Windthrow Severity
Landsat 5 TM images processed with LEDAPS (Landsat Ecosystem Disturbance Adaptive Processing System) [32] were used to detect the occurrence of windthrows. For Oroza, Landsat image P063R03 from 17 April 1988 (prior to disturbance) and 24 September 1988 (after disturbance) were used; for Napo, P062R03 from 19 August 1998 and 22 October 1998; and for Nauta, P063R03 from 2 September 2009 and 7 December 2009. Landsat 5 imagery with Ecosystem Disturbance Adaptive Processing System (LEDAPS) is available in the Google Earth Engine https://earthengine.google.com,(accessed on 24 February 2020) [33], which is the platform used for our imagery analysis and Landsat imagery processing. Windthrows were identified by their spectral characteristics (endmembers) and their distinctive shape diverging from a central area with radiant corridors separated by forest [1]. We used spectral mixture analysis (SMA) [34] to quantify the fraction of distinct endmembers. Specifically, we used image-derived endmembers: green photosynthetic vegetation (GV), non-photosynthetic vegetation (NPV), and shadows. The fractions of NPV and GV were then normalized without shadows as NPV/(GV + NPV) and GV/(GV

Assessing Windthrow Severity
Landsat 5 TM images processed with LEDAPS (Landsat Ecosystem Disturbance Adaptive Processing System) [32] [33], which is the platform used for our imagery analysis and Landsat imagery processing. Windthrows were identified by their spectral characteristics (endmembers) and their distinctive shape diverging from a central area with radiant corridors separated by forest [1]. We used spectral mixture analysis (SMA) [34] to quantify the fraction of distinct endmembers. Specifically, we used image-derived endmembers: green photosynthetic vegetation (GV), non-photosynthetic vegetation (NPV), and shadows. The fractions of NPV and GV were then normalized without shadows as NPV/(GV + NPV) and GV/(GV + NPV) [3,30,34]. Metrics of the windthrow area and damage are derived from ∆NPV [3,6,15,30,35]. Using the ∆NPV, we estimated the percentage of trees directly toppled or killed in the studied windthrows by using a locally adjusted predictive windthrow tree-mortality model (Equation (1)) [12]: To assign the mortality value by subplot (10 × 30 m), we resampled the image at a greater spatial resolution (from 30 × 30 m to 3 × 3 m), and extracted weighted ∆NPV values based on the pixels comprising subplots using the Zonal statistics tool available in QGIS 3.10.3. The ∆NPV values were used to classify windthrow severity at the subplot level, complementing previous studies in the CA [8].

Forest Inventory
At each site, we installed 10 m wide transects of varying lengths, with a total area of three hectares for Napo and Oroza, and 3.09 ha for Nauta. In Napo and Oroza, the transects were subdivided into 100 subplots of 300 m 2 each (i.e., 30 m × 10 m), and 103 subplots for Nauta, yielding a total of 303 subplots for our analysis. The transects encompassed the entire gradient of windthrow tree-mortality derived from our remote-sensing metric [6,12]. In each subplot we measured the diameter of trees > 10 cm DBH (diameter at breast height, 1.3 m). Combining all sites, we recorded 4889 living trees. A botanical sample was collected from each recorded species for tree identification. A summary of forest floristics characteristics by severity is shown in Appendix A. Tree locations were determined using a local reference system (UTM 18S) and the following procedure. First, a handheld GPS receiver (Global Position System, Garmin Map76 CSx, Olathe, Kansas, U.S.A) was used to locate the start point of each transect, then the "x" and "y" coordinates of each tree were measured using a rangefinder laser (Trimble Laserace 1000, Sunnyvale, CA, USA) based on their distance and angle from the transect's starting position.

Data Processing and Analysis
The basal area (BA) of recorded trees m 2 ha −1 was estimated from the diameter at breast height assuming a circular trunk shape [36]. As there are no locally adjusted allometries, we estimated individual tree aboveground biomass (AGB) using references from other Amazon regions (Models 1, 2 and 3, Table 1). The AGB at the subplot level was calculated by summing the individual biomass of trees recorded in each subplot. Since Models 1 and 2 include wood density as a predictor, we compiled wood density values of recorded species from previous studies conducted in the Amazon Basin [5,8,38]. Values were assigned using data at the species, genus, or family level (in order of preference), depending on the availability and geographical proximity of our study region. Model 3 was applied to individual trees from species for which wood density was not reported.
Forest structural attributes such as biomass can be over-and/or underestimated due to spatial variations in small plots containing relatively large individual trees [5,6,39]. Previous studies from the Central Amazon suggest that robust estimates of biomass require sample units with areas between 0.08 ha to 0.12 ha [8,39]. Therefore, to reduce intrinsic spatial variations in biomass estimates, we binned subplots from our study sites based on their windthrow severity. The final area of our binned plots ranged between 0.09-0.12 ha.

Statistical Analysis
A chi-square test at the 5% significance level was used to compare the diameter distribution of trees in old-growth plots. A Student's t-test was used to assess whether differences in mean DBH, basal area, and average AGB were statistically significant between disturbed and old-growth forest. To compare tree size, basal area, and AGB among successional stages, we used a one-way ANOVA. For reporting forest recovery, we assumed a 10% uncertainty around the respective undisturbed reference, i.e., a subplot was considered 'recovered' if it attained 90% of the old-growth value for a given attribute. We estimated the time required to recover old-growth levels of basal area and AGB by fitting linear regressions to the subplot-level data using the time since windthrow as the predictor. We further calculated 90% confidence bands of the response in order to visualize uncertainties. Although our high number of subplots (total of 303) allows for calculating confidence intervals and p-values independent of data distribution, these may be overly optimistic as they do not account for site-level autocorrelation. The statistical analyses were carried out in the statistical software package R 4.0.3 [40].

Subplot Distribution by Disturbance Severity Class
Estimates of windthrow tree-mortality (Equation (1)) and the locations of transects at respective sites allowed us to calculate the number of plots and the area occupied by each severity class (Table 2). These classes encompass the range of disturbance severity from old-growth forest to low, moderate, high, and extreme severity, and are given in Table 2.

Structural Attributes of Old-Growth Forests
Old-growth forest subplots at our study sites had a mean (± standard deviation) trees per hectare (TPH) of 562 ± 87 trees ha −1 , BA of 24.8 ± 2.3 m 2 ha −1 and AGB of 380 ± 156 Mg ha −1 . We assessed possible pre-disturbance differences among old-growth forests by assessing the mean diameter of trees, the basal area and the AGB across oldgrowth plots. The results of the ANOVA indicated a significant difference in forest structure among the study sites for BA (p = 0.003) and AGB (p = 0.002), but no significant differences for mean DBH (p = 0.079). A post-hoc T-test (p-adjust using the Benjamin-Hochberg method) indicated that there is no statistical evidence of differences in mean DBH (p = 0.31), BA (p = 0.55), and AGB (p = 0.71) at the Napo and Nauta sites. The Oroza site showed significant differences from the Napo site for BA (p = 0.02) but no difference from the Nauta site (p = 0.06). For AGB, Oroza was significantly different from Nauta and Napo (p = 0.007 and p = 0.004, respectively). For mean DBH, Oroza did not differ significantly from Napo (p = 0.07) or Nauta (p = 0.3). We did not find significant differences in the size distribution of trees recorded at varying windthrow severities. However, mean DBH was significantly reduced in extremely disturbed plots 2 years after windthrow (t-test p = 0.03), but differences were not significant 12 years (t-test p = 0.1) and 22 years (t-test p = 0.16) after windthrow. We therefore suggest that in extremely disturbed areas, DBH has recovered to old-growth levels by year 12. For moderate and high severities, no significant differences were observed 22 years after windthrow, with t-test values of p = 0.58 and p = 0.50, respectively.
The three sites, spanning 2 to 22 years after disturbance, show clear recovery of structural attributes in the decades following the respective windthrow events. Figure 2a shows the number of trees per hectare as a function of windthrow severity and recovery time. Two years after windthrow disturbance, the number of trees per hectare was 382 (± 35) trees ha −1 in extremely disturbed areas. This rapidly increased to 599 (± 55) trees ha −1 and 578 (±112) trees ha −1 12 and 22 years after windthrow disturbance, respectively. This apparently fast recovery of TPH was also observed for high and moderate windthrow severities. Figure 2b shows the basal area (BA) as a function of windthrow severity and recovery time. Comparing the BA of extremely disturbed forest 2 years following the windthrow event (12.46 ± 4.5 m 2 ha −1 ) with that of extremely disturbed forest after 12 years (23.52 ± 4.51 m 2 ha −1 ) and 22 years (24.46 ± 5.03 m 2 ha −1 ) yields a gradual increase over time. Similar recovery patterns are observed for the high, moderate and low severity categories. 0.004, respectively). For mean DBH, Oroza did not differ significantly from Napo (p = 0.07) or Nauta (p = 0.3).

Mean Diameter, Tree Density and Basal Area
We did not find significant differences in the size distribution of trees recorded at varying windthrow severities. However, mean DBH was significantly reduced in extremely disturbed plots 2 years after windthrow (t-test p = 0.03), but differences were not significant 12 years (t-test p = 0.1) and 22 years (t-test p = 0.16) after windthrow. We therefore suggest that in extremely disturbed areas, DBH has recovered to old-growth levels by year 12. For moderate and high severities, no significant differences were observed 22 years after windthrow, with t-test values of p = 0.58 and p = 0.50, respectively.
The three sites, spanning 2 to 22 years after disturbance, show clear recovery of structural attributes in the decades following the respective windthrow events. Figure 2a shows the number of trees per hectare as a function of windthrow severity and recovery time. Two years after windthrow disturbance, the number of trees per hectare was 382 (± 35) trees ha −1 in extremely disturbed areas. This rapidly increased to 599 (± 55) trees ha −1 and 578 (±112) trees ha −1 12 and 22 years after windthrow disturbance, respectively. This apparently fast recovery of TPH was also observed for high and moderate windthrow severities. Figure 2b shows the basal area (BA) as a function of windthrow severity and recovery time. Comparing the BA of extremely disturbed forest 2 years following the windthrow event (12.46 ± 4.5 m 2 ha −1 ) with that of extremely disturbed forest after 12 years (23.52 ± 4.51 m 2 ha −1 ) and 22 years (24.46 ± 5.03 m 2 ha −1 ) yields a gradual increase over time. Similar recovery patterns are observed for the high, moderate and low severity categories.

Aboveground Tree Biomass
Out of the 4889 trees recorded in our study, 2171 (44.5%) were assigned wood density values at the species level, 1554 (31.7%) at the genus level and 523 (10.7%) at the family level. We did not assign wood density values for 641 (13.1%) trees. We used 95 binned plots distributed over 4 severities and 3 measured years (ranging from 0.09 ha. to 0.12 ha.). Estimates of AGB differed when using Model 1 [37] versus Model 2 [5], mainly due to the

Aboveground Tree Biomass
Out of the 4889 trees recorded in our study, 2171 (44.5%) were assigned wood density values at the species level, 1554 (31.7%) at the genus level and 523 (10.7%) at the family level. We did not assign wood density values for 641 (13.1%) trees. We used 95 binned plots distributed over 4 severities and 3 measured years (ranging from 0.09 ha. to 0.12 ha.). Estimates of AGB differed when using Model 1 [37] versus Model 2 [5], mainly due to the relatively higher values given for trees larger than 50 cm DBH by Model 1. However, regardless of which model was used, the overall trajectory of AGB recovery did not change (Appendix B).
Extreme disturbance (>60% mortality) reduced the mean AGB to 163 ± 68 Mg ha −1 (2 years after windthrow). AGB increased to 323.4 ± 138.7 Mg ha −1 and 327.5 ± 86.2 Mg ha −1 , 12 and 22 years after disturbance, respectively. This AGB recovery process was also observed for sites with high severity. Although AGB loss was considerably lower at low and moderate severity, these areas showed an increased biomass compared to old-growth levels 12 years after windthrow. AGB values differed among studied old-growth areas (see details in Appendix C). For these reasons, AGB recovery in disturbed areas was analyzed relative to old-growth values. Thus, a relative sequence of AGB recovery was made ( Figure 3A, bottom panel). As suggested by a linear model (Figure 4-Iquitos), AGB stocks at all windthrow severities will recover to~90% of that observed in contiguous old-growth forest over a time span of 22 years. relatively higher values given for trees larger than 50 cm DBH by Model 1. However, regardless of which model was used, the overall trajectory of AGB recovery did not change (Appendix B). Extreme disturbance (>60% mortality) reduced the mean AGB to 163 ± 68 Mg ha −1 (2 years after windthrow). AGB increased to 323.4 ± 138.7 Mg ha −1 and 327.5 ± 86.2 Mg ha −1 , 12 and 22 years after disturbance, respectively. This AGB recovery process was also observed for sites with high severity. Although AGB loss was considerably lower at low and moderate severity, these areas showed an increased biomass compared to old-growth levels 12 years after windthrow. AGB values differed among studied old-growth areas (see details in Appendix C). For these reasons, AGB recovery in disturbed areas was analyzed relative to old-growth values. Thus, a relative sequence of AGB recovery was made (Figure 3a, bottom panel). As suggested by a linear model (Figure 4-Iquitos), AGB stocks at all windthrow severities will recover to ~90% of that observed in contiguous old-growth forest over a time span of 22 years.

Patterns of Recovery after Windthrow
Differences in forest attributes exist across the Amazon Basin [41,42], including treemortality predictors [10,43] that can explain different mortality/survival mechanisms when compared across a large region. Winds are an important mechanism of natural disturbance in forests worldwide [44][45][46]. Windthrows produce varying levels of tree damage/mortality that change forest structure [6,20] and can initiate varying successional trajectories [8,15,28]. Our results give further evidence to the importance of wind as a major disturbance mechanism in Amazon forests [9,12,15,47] Forest inventories at windthrow sites in the Northwestern Amazon (our plots) and Central Amazon [5,8,15] followed the same overall sampling and analysis design. The degree of windthrow tree mortality estimated from ∆NPV was higher (up to 95%) in our Northwestern Amazon sites compared to the Central Amazon (up to 70%) [5,12,15]. For that reason, we included one additional severity class to represent extreme windthrow tree-mortality (>60%).
The 562 ± 87 trees ha −1 in our old-growth forests is similar to the number reported by other studies conducted in the Northwestern Amazon, for example 608-616 trees ha −1 [48], 559 ± 74 trees ha −1 [41], 589 ± 20 trees ha −1 [49], as well as in the Central Amazon (583 ± 46 trees ha −1 ) [15]. Our data on stem density ( Figure 3A, top panel) shows that 12 years after windthrow, the windthrown forest reaches the same values observed in the adjacent/contiguous old-growth, and maintains this pattern at least until 22 years of recovery. Compared to plots in the Central Amazon, (Figure 3B, top panel) the patterns of forest recovery are similar in terms of TPH and the time required to recover to old-growth values for both regions is between 12 and 14 years, independent of the severity. Tree density was smaller in the Northwestern Amazon indicating that, for a given windthrow severity, mortality was relatively higher in this region than in the Central Amazon. Tree density recovery following disturbances was also faster in Iquitos (NWA) compared to Manaus (CA). In the Northwestern Amazon, tree density was reduced to 377 ± 89 trees ha −1 and 382 ± 35 trees ha −1 2 years after windthrow disturbance in the >40% and > 60% tree-mortality categories, respectively. In the Central Amazon, tree density was 422 ± 121 trees ha −1 4 years after disturbance in the >40% tree-mortality category [15].

Patterns of Recovery after Windthrow
Differences in forest attributes exist across the Amazon Basin [41,42], including treemortality predictors [10,43] that can explain different mortality/survival mechanisms when compared across a large region. Winds are an important mechanism of natural disturbance in forests worldwide [44][45][46]. Windthrows produce varying levels of tree damage/mortality that change forest structure [6,20] and can initiate varying successional trajectories [8,15,28]. Our results give further evidence to the importance of wind as a major disturbance mechanism in Amazon forests [9,12,15,47] Forest inventories at windthrow sites in the Northwestern Amazon (our plots) and Central Amazon [5,8,15] followed the same overall sampling and analysis design. The degree of windthrow tree mortality estimated from ∆NPV was higher (up to 95%) in our Northwestern Amazon sites compared to the Central Amazon (up to 70%) [5,12,15]. For that reason, we included one additional severity class to represent extreme windthrow tree-mortality (>60%).
The 562 ± 87 trees ha −1 in our old-growth forests is similar to the number reported by other studies conducted in the Northwestern Amazon, for example 608-616 trees ha −1 [48], 559 ± 74 trees ha −1 [41], 589 ± 20 trees ha −1 [49], as well as in the Central Amazon (583 ± 46 trees ha −1 ) [15]. Our data on stem density ( Figure 3A, top panel) shows that 12 years after windthrow, the windthrown forest reaches the same values observed in the adjacent/contiguous old-growth, and maintains this pattern at least until 22 years of recovery. Compared to plots in the Central Amazon, (Figure 3B, top panel) the patterns of forest recovery are similar in terms of TPH and the time required to recover to old-growth values for both regions is between 12 and 14 years, independent of the severity. Tree density was smaller in the Northwestern Amazon indicating that, for a given windthrow severity, mortality was relatively higher in this region than in the Central Amazon. Tree density recovery following disturbances was also faster in Iquitos (NWA) compared to Manaus (CA). In the Northwestern Amazon, tree density was reduced to 377 ± 89 trees ha −1 and 382 ± 35 trees ha −1 2 years after windthrow disturbance in the >40% and > 60% tree-mortality categories, respectively. In the Central Amazon, tree density was 422 ± 121 trees ha −1 4 years after disturbance in the >40% tree-mortality category [15].
Basal area in old-growth forests in the Northwestern Amazon was previously reported as~28 m 2 ha −1 [41] and 27.65 (±2.0) m 2 ha −1 [49]. These values are similar to those found at our study sites in the Northwestern Amazon (24.8 ± 2.3 m 2 ha −1 ) and those in the Central Amazon (26.8 ± 2.4 m 2 ha −1 ) [15]. In our study, the estimated recovery time for the relative basal area ( Figure 3A, middle panel) shows a gradual increase in all the windthrow severities. Compared with the Central Amazon ( Figure 3B, top and middle panels), we observed a faster recovery of tree density and basal area in the Northwestern Amazon. This also contributes to a more rapid recovery of biomass [15,50] and may be related to changes in floristic composition and species diversity [15,51].

Time Span of Biomass Recovery
In tropical forests, the recovery of biomass following disturbance can vary greatly depending on the latitude, type and severity of the disturbance. In secondary forests created by human disturbances (e.g., slash-and-burn, and clear cut), biomass recovery to old-growth conditions can take from 66 [52] to >100 years [31,[53][54][55]. These differences in time reflect, among other factors, the availability of nutrients for regrowth [56]. A hurricane in Puerto Rico reduced biomass by 50% [57]. Moreover, impacts and recovery times for this disturbance depended strongly on the degree of hurricane damage, with biomass recovery taking from 4 to 6 years (F0 damage), and from 50 to 150 years (F3 damage) [58,59]. This pattern was also observed after selective logging in Tapajos, Brazil (Central Amazon). Biomass recovery took~13 years [60] and minimally affected forest carbon and water exchanges [61].
Previous studies in old-growth Amazon forests reported AGB ranging from 303 to 385 Mg ha −1 [62], 155 to 425 Mg ha −1 [63] and~256 Mg ha −1 [64]. Regional-scale variations in biomass are well documented and likely related to factors such as climate [21,22], soils and nutrient availability, or geomorphological characteristics [23]. The causes of regional variations in vegetation structure across different forest types [64] and the main drivers of tree mortality are not yet fully understood [10]. Our results highlight the importance of large-scale windthrows in shaping forest structure in the Northwestern Amazon.
In the Amazon, windthrow distribution extends from the Northwestern Amazon to the Central Amazon [12], showing a spatial and temporal variability in gap sizes [11,65] and mortality rates [6,8,9,12]. In our study, the estimated time to recover AGB increased slightly with windthrow severity. In our study, the recovery to old-growth levels occurred after 22 years ( Figure 3A, bottom panel), whereas in the Central Amazon, 24 years were only enough to recover the old-growth levels of attributes observed in forests exposed to low-severity windthrow, and showed a slower growth increment for moderate and high severities even 27 years after windthrow ( Figure 3B, bottom panel).
We described the trajectories of biomass recovery in the Northwestern and Central Amazon using linear models (Figure 4). For the Northwestern Amazon, we estimated recovery to be at least > 90% of the old-growth condition after 14 years, 15 years, 18 years, and 20 years since windthrow for low, moderate, extreme, and high windthrow treemortality, respectively. For the Central Amazon we estimated~10 years,~30 years, and 37 years for low, high, and moderate windthrow tree-mortality, respectively (details in Supplementary Material SM1, SM2). Differences in the width of confident intervals between the Northwestern and Central Amazon in Figure 4 reflect the number of binned plots used for each site and the numbers of measured years. In the Central Amazon, more complex models such as the generalized additive model (GAM) were used for estimating how biomass recovered to pre-disturbance levels (from 27 to 40 years, maximum) [8].
Our study indicated a faster recovery of biomass in the Northwestern Amazon than in the Central Amazon, which may be explained by the higher productivity in this region [26], and potentially related geographic variations in community attributes, including larger tree diversity [42] which can influence tree mortality patterns [12]. In addition, high competition in combination with water availability [25,66] promotes high productivity, and species that sacrifice defenses in favor of the competitive advantage obtained by rapid growth and low wood density may be favored in regions with severe disturbance regimes [43]. Thus, faster recovery could be the result of adaptation to more frequent windthrow disturbance in the Northwestern Amazon [12], which provides supporting evidence for the growth-defense hypothesis [67].
Ecologists have long recognized that disturbances and recovery processes overlap in spatial and temporal dimensions [68]. In our study sites, the density and size distribution of trees had returned to old-growth levels only 12 years after windthrow. Furthermore, basal area and biomass recovered to pre-disturbance levels within 20 years of the disturbance. This is~50% faster than the recovery from windthrows observed in the Central Amazon. Future research on forest recovery from windthrows should include a component of geospatial analysis to help explain individual tree response [8].

Conclusions
This study demonstrated rapid recovery of tree density, basal area, and AGB in the two decades following windthrow events that caused mortality rates >60% in forests near Iquitos, Peru. The observed recovery rates are nearly twice as fast as those reported for a comparable study conducted in the slower-growing forests of the Central Amazon. The variations in forest damage and the recovery dynamics across this geographical and climatic gradient emphasize the effect that extreme wind and rain associated with convective storms can have on the spatial variability of forest structure across the Amazon Basin.