Windthrow Variability in Central Amazonia

: Windthrows are a recurrent disturbance in Amazonia and are an important driver of forest dynamics and carbon storage. In this study, we present for the ﬁrst time the seasonal and interannual variability of windthrows, focusing on Central Amazonia, and discuss the potential meteorological factors associated with this variability. Landsat images over the 1998–2010 time period were used to detect the occurrence of windthrows, which were identiﬁed based on their spectral characteristics and shape. Here, we found that windthrows occurred every year but were more frequent between September and February. Organized convective activity associated with multicell storms embedded in mesoscale convective systems, such as northerly squall lines (that move from northeast to southwest) and southerly squall lines (that move from southwest to northeast) can cause windthrows. We also found that southerly squall lines occurred more frequently than their previously reported ~50 year interval. At the interannual scale, we did not ﬁnd an association between El Niño-Southern Oscillation (ENSO) and windthrows.


Introduction
Windthrows are a recurrent form of tree mortality in the Amazon. They are produced by downbursts [1][2][3], which are strong descending winds associated with severe convective storms [1][2][3][4][5] that create gaps of uprooted or broken trees [6]. These gaps vary in size from a single tree to thousands of hectares of forest [4,[6][7][8]. Windthrows affect the residence time of woody biomass, which, in turn, affects patterns of productivity and biomass [8,9], floristic composition [10][11][12], and soil composition [13] in the basin. Recent studies using demographic models have shown that an increase in windthrow frequency promotes a decrease in biomass and leaf area index in Central Amazonia [14]. These results suggests that windthrows may have cascading effects and represent an important and overlooked source of uncertainty in climate predictions, especially given more intense rainfall events are expected to occur in the future [15]. Though efforts have been made to understand 1.
What is the seasonal and interannual variability of windthrows? 2.
Are windthrows associated with ENSO?
The temporal variability of windthrows in the Amazon is currently unknown. This study therefore represents the first record of windthrow variability in the Amazon. As a pioneering study, we focus on the generalities of atmospheric events responsible for the variability of windthrows.

Meteorological Data
We characterized the current climatology in Central Amazonia based on the period 1971-2000 [39] using the Global Precipitation Climatological Centre data Reanalysis version 7 for rainfall [40] and the Climatic Research Unit Time series v3.23 data [41] for temperature. Both data sets are available at 0.5° horizontal resolution on a monthly basis. Our study area is characterized by a mean annual temperature of 27 °C and mean annual rainfall of 2365 mm with the dry season (rainfall <100 mm·month −1 , [42]) falling between July and September ( Figure 1).
The Tropical Rainfall Measuring Mission (TRMM) data [43] version 7 were also used in this study at monthly (3B43) and 3 h temporal resolution (3B42) from 1998 to 2010. These gridded data are at a 0.25° × 0.25° spatial resolution from 50° South to 50° North latitude. The 3B42 estimates are a combination of microwave and infrared precipitation estimates calibrated and merged with

Meteorological Data
We characterized the current climatology in Central Amazonia based on the period 1971-2000 [39] using the Global Precipitation Climatological Centre data Reanalysis version 7 for rainfall [40] and the Climatic Research Unit Time series v3.23 data [41] for temperature. Both data sets are available at 0.5 • horizontal resolution on a monthly basis. Our study area is characterized by a mean annual temperature of 27 • C and mean annual rainfall of 2365 mm with the dry season (rainfall <100 mm·month −1 , [42]) falling between July and September ( Figure 1).
The Tropical Rainfall Measuring Mission (TRMM) data [43] version 7 were also used in this study at monthly (3B43) and 3 h temporal resolution (3B42) from 1998 to 2010. These gridded data are at a 0.25 • × 0.25 • spatial resolution from 50 • South to 50 • North latitude. The 3B42 estimates are a combination of microwave and infrared precipitation estimates calibrated and merged with observations, which are then rescaled to a monthly time step to produce the 3B43 data [43]. TRMM 3B43 and TRMM 3B42 will be referred to hereafter as TRMMmo and TRMM3h, respectively.
We also used reanalysis data from the National Centers for Environmental Prediction (NCEP) [44] to study the SSLs.

Windthrow Identification
The long time series of Landsat imagery (30 m horizontal resolution) is appropriate to study the temporal variability of windthrows. Thematic Mapper images from Landsat 5 and Landsat 7 satellites (L5 and L7, respectively) are available in the archive of the Google Earth Engine (GEE) [45]. Fifty-nine images from September 1998 through August 2010 were analyzed in this study. Most of the images used had less than 30% cloud cover and images from June to November were the most common images analyzed (Figure 2), a pattern that coincides with the beginning and end of the dry season ( Figure 1) and therefore less cloud cover. Figure 2 also shows the difficulty in analyzing the traditional four seasons (DJF, MAM, JJA, SON) in Central Amazonia using Landsat imagery. We have therefore grouped images from the rainy (SONDJF) and dry (MAMJJA) seasons. Although there may be some uncertainty surrounding the seasonal assignment of a windthrow, the occurrence of the windthrow is clearly identified in all cases. observations, which are then rescaled to a monthly time step to produce the 3B43 data [43]. TRMM 3B43 and TRMM 3B42 will be referred to hereafter as TRMMmo and TRMM3h, respectively. We also used reanalysis data from the National Centers for Environmental Prediction (NCEP) [44] to study the SSLs.

Windthrow Identification
The long time series of Landsat imagery (30 m horizontal resolution) is appropriate to study the temporal variability of windthrows. Thematic Mapper images from Landsat 5 and Landsat 7 satellites (L5 and L7, respectively) are available in the archive of the Google Earth Engine (GEE) [45]. Fifty-nine images from September 1998 through August 2010 were analyzed in this study. Most of the images used had less than 30% cloud cover and images from June to November were the most common images analyzed (Figure 2), a pattern that coincides with the beginning and end of the dry season ( Figure 1) and therefore less cloud cover. Figure 2 also shows the difficulty in analyzing the traditional four seasons (DJF, MAM, JJA, SON) in Central Amazonia using Landsat imagery. We have therefore grouped images from the rainy (SONDJF) and dry (MAMJJA) seasons. Although there may be some uncertainty surrounding the seasonal assignment of a windthrow, the occurrence of the windthrow is clearly identified in all cases. To identify windthrows, a spectral mixture analysis (SMA) [46,47] was applied to the Landsat images. SMA quantifies the per pixel fraction of endmembers which sums to match the full pixel spectrum of the image [46]. Image-derived endmembers of green (photosynthetic) vegetation (GV), non-photosynthetic vegetation (NPV), and shade were used. A shade endmember was included to account for effects related to view angle, topography, shading, and shadows from clouds. Endmembers for GV was green cecropia and for shade a lake in our Landsat scene as shown in Chapter 7 in Adams and Gillespie [48]. In windthrow areas, a large increase in bare wood from downed trees is exposed to the satellite sensor, yielding a high mid-infrared reflectance (band 5, 1550-1750 nm, in Landsat 5 and Landsat 7) that lasts for about a year [3,4]. The fractions of GV and NPV were then normalized without shade [48] as GV/(GV + NPV) and NPV/(GV + NPV). Our goal was to detect windthrows, and applying SMA to the DN (digital number) of the Landsat data was reasonable given SMA in Amazon forests does not produce significantly different results using DN or atmospherically corrected images [48,49]. A windthrow area was determined from the polygon containing the windthrow using available tools in the GEE platform. Only windthrows located >2 km away from urban areas and larger than 5 ha were considered in our analysis. Examples of windthrows over our study area have been published in our previous studies (e.g., Figure 2 in [3] and Figure S5 in [8]). To identify windthrows, a spectral mixture analysis (SMA) [46,47] was applied to the Landsat images. SMA quantifies the per pixel fraction of endmembers which sums to match the full pixel spectrum of the image [46]. Image-derived endmembers of green (photosynthetic) vegetation (GV), non-photosynthetic vegetation (NPV), and shade were used. A shade endmember was included to account for effects related to view angle, topography, shading, and shadows from clouds. Endmembers for GV was green cecropia and for shade a lake in our Landsat scene as shown in Chapter 7 in Adams and Gillespie [48]. In windthrow areas, a large increase in bare wood from downed trees is exposed to the satellite sensor, yielding a high mid-infrared reflectance (band 5, 1550-1750 nm, in Landsat 5 and Landsat 7) that lasts for about a year [3,4]. The fractions of GV and NPV were then normalized without shade [48] as GV/(GV + NPV) and NPV/(GV + NPV). Our goal was to detect windthrows, and applying SMA to the DN (digital number) of the Landsat data was reasonable given SMA in Amazon forests does not produce significantly different results using DN or atmospherically corrected images [48,49]. A windthrow area was determined from the polygon containing the windthrow using available tools in the GEE platform. Only windthrows located >2 km away from urban areas and larger than 5 ha were considered in our analysis. Examples of windthrows over our study area have been published in our previous studies (e.g., Figure 2 in [3] and Figure S5 in [8]).

Dating of Windthrows
Windthrows were identified by their fan-shaped form [4] and high NPV values, as we have shown in previous studies [3,7,8]. We visually inspected each image that preceded a windthrow image to ensure that the prior image showed undisturbed forest. We chose to study the seasonal and annual variability of windthrows because of the availability of Landsat images and the variability of rainfall over these time periods.
Annually. We assigned the timing of windthrows to the hydrological year (HY, September to August) in which they most likely occurred. We used the date of the Landsat image in which a windthrow first appeared to identify the windthrow date, although the windthrow could have occurred at any time between the previous and current Landsat image, introducing uncertainty in the date assignment. Seasonally. To study windthrow seasonality, we examined rainy season (SONDJF, months are identified by their first letter) and dry season (MAMJJA) windthrow frequency. A larger dating uncertainty occurred with rainy season (SONDJF) images because of cloud cover and image availability over this time period. On average, 2 ± 1 (mean ± standard deviation) windthrows could have occurred in the previous season (MAMJJA), except in the hydrological years 2001-2002, 2003-2004, 2005-2006, 2006-2007, and 2007-2008 when 12 ± 3 windthrows that could have occurred in the previous season. A shorter time scale (e.g., monthly rather than seasonal) is difficult to analyze due to image availability and an abundance of cloud cover.
Another source of uncertainty that makes windthrow dating difficult is the fact that the edge overlap of the Landsat tiles varies with time. If a windthrow for a given year appeared on the border of an image, but this same border did not appear in subsequent images then that windthrow was excluded from our analysis.

ENSO Years
Sea surface temperature (SST) data [50] is used to examine anomalies in the Pacific region (5 • N-5 • S, 120 • W-170 • W, the Niño 3.4 region) related to ENSO (positive values indicate warm El Niño periods and negative values indicate cold La Niña periods). A SST deviation ≥0.5 • C is used by the Climatic Prediction Center/National Oceanic and Atmospheric Administration to define the occurrence of an El Niño or La Niña. The list of ENSO years as well as further details are available at [51]. In this study, we used this list to identify ENSO years in our time series. Figure 3 shows the occurrence of windthrows over the study area from 1998 to 2010. A few windthrows appeared within the black background of the image; these were related to changes in the Landsat tile cover areas. Populated areas were not considered in our analysis and therefore no windthrows were mapped in these areas. For instance, the southeastern corner of the Landsat image has high anthropogenic activity and therefore few windthrows were mapped here. The histogram of windthrow sizes was skewed to the left indicating that the smallest windthrows were the most frequent, which is consistent with previous studies [7,8]. Windthrows in bins (data in intervals) of 20, 40, and 60 ha represented 42%, 28%, and 12% of windthrows events, respectively. was selected due to its free cloud cover condition) is shown for spatial context: blue represents water bodies, pink and light green represent anthropogenic areas, and dark green represents old-growth forest. Squares represent the centroid of windthrows with the color representing the year of their occurrence as defined in the figure legend.

Results
Our results showed ( Figure 4) a high seasonal variability of windthrows, but, in general, the rainy season (SONDJF) had a greater number of windthrows than the dry season (MAMJJA). This pattern is similar for the HYb case (results not shown). For the MAMJJA 2002-2003 and SONDJF 2008-2009 seasons, the number of events dated was zero. We found that 75% of windthrows occurred in the rainy season and 25% in the dry season. We also considered the seasonal dating uncertainty when a given windthrow could have occurred in the previous season. If we account for these cases, only 59% of windthrows occurred during the rainy season. The seasonal occurrence of windthrows observed in our data agrees with previous studies that have dated windthrows [3,16]. Abundant cloud cover made it difficult to date windthrows during the rainy season (SONDJF), particularly during the 2007-2008 La Niña year when 18 windthrows could have occurred in the previous season. Our results showed ( Figure 4) a high seasonal variability of windthrows, but, in general, the rainy season (SONDJF) had a greater number of windthrows than the dry season (MAMJJA). This pattern is similar for the HYb case (results not shown). For the MAMJJA 2002-2003and SONDJF 2008 seasons, the number of events dated was zero. We found that 75% of windthrows occurred in the rainy season and 25% in the dry season. We also considered the seasonal dating uncertainty when a given windthrow could have occurred in the previous season. If we account for these cases, only 59% of windthrows occurred during the rainy season. The seasonal occurrence of windthrows observed in our data agrees with previous studies that have dated windthrows [3,16]. Abundant cloud cover made it difficult to date windthrows during the rainy season (SONDJF), particularly during the 2007-2008 La Niña year when 18 windthrows could have occurred in the previous season. The number of windthrows that occur annually mostly covaries with annual rainfall. However, a few key mismatches ( Figure 5) prevent a high coefficient of determination (r 2 < 0.1 p < 0.001) between windthrows and rainfall for both the HY and HYb cases. HY2004-2005 and HY2008-2009 presented the lowest [39] and highest [52] rainfall amounts, respectively (the pattern was similar for the HYb case), and, rather counterintuitively, these years had the highest (HY2004-2005) and lowest (HY2008-2009) occurrence of windthrows. When these two years were omitted, the data showed a much stronger relationship between rainfall and windthrows in the HY case (r 2 = 0.7, p < 0.001). Interestingly, this relationship was not captured in the HYb case (r 2 < 0.1), revealing the importance of properly dating windthrows. On average, the annual number of studied windthrows was 16 ± 10 events (mean ± SD). In the HY2004-2005, a large number of windthrows were observed even though that year contained the lowest rainfall amount. In this year, the large occurrence of windthrows was produced The number of windthrows that occur annually mostly covaries with annual rainfall. However, a few key mismatches ( Figure 5) prevent a high coefficient of determination (r 2 < 0.1 p < 0.001) between windthrows and rainfall for both the HY and HYb cases. HY2004-2005 and HY2008-2009 presented the lowest [39] and highest [52] rainfall amounts, respectively (the pattern was similar for the HYb case), and, rather counterintuitively, these years had the highest (HY2004-2005) and lowest (HY2008-2009) occurrence of windthrows. When these two years were omitted, the data showed a much stronger relationship between rainfall and windthrows in the HY case (r 2 = 0.7, p < 0.001). Interestingly, this relationship was not captured in the HYb case (r 2 < 0.1), revealing the importance of properly dating windthrows. On average, the annual number of studied windthrows was 16 ± 10 events (mean ± SD). The number of windthrows that occur annually mostly covaries with annual rainfall. However, a few key mismatches ( Figure 5) prevent a high coefficient of determination (r 2 < 0.1 p < 0.001) between windthrows and rainfall for both the HY and HYb cases. HY2004-2005 and HY2008-2009 presented the lowest [39] and highest [52] rainfall amounts, respectively (the pattern was similar for the HYb case), and, rather counterintuitively, these years had the highest (HY2004-2005) and lowest (HY2008-2009) occurrence of windthrows. When these two years were omitted, the data showed a much stronger relationship between rainfall and windthrows in the HY case (r 2 = 0.7, p < 0.001). Interestingly, this relationship was not captured in the HYb case (r 2 < 0.1), revealing the importance of properly dating windthrows. On average, the annual number of studied windthrows was 16 ± 10 events (mean ± SD). In the HY2004-2005, a large number of windthrows were observed even though that year contained the lowest rainfall amount. In this year, the large occurrence of windthrows was produced In the HY2004-2005, a large number of windthrows were observed even though that year contained the lowest rainfall amount. In this year, the large occurrence of windthrows was produced by a SSL that crossed the Amazon basin in January of 2005 producing a large number of windthrows in the Manaus region [3]. However, only four NSLs [17] were recorded that month, the smallest number recorded during any month in 2005 [53].
We identified several SSLs (November and December 1998, January and February 2002, January 2004 and January 2005) using the TRMM3h data ( Figure 6). Interestingly, these years also contain a high number of windthrows, particularly HY1998-1999, a La Niña year. A peak in windthrow frequency was also observed during HY1999-2000, another La Niña year. The synoptic features associated with the November 1998 SSL event are shown in Figures 7 and 8 by a SSL that crossed the Amazon basin in January of 2005 producing a large number of windthrows in the Manaus region [3]. However, only four NSLs [17] were recorded that month, the smallest number recorded during any month in 2005 [53]. We identified several SSLs (November and December 1998, January and February 2002, January 2004 and January 2005) using the TRMM3h data ( Figure 6). Interestingly, these years also contain a high number of windthrows, particularly HY1998-1999, a La Niña year. A peak in windthrow frequency was also observed during HY1999-2000, another La Niña year. The synoptic features associated with the November 1998 SSL event are shown in Figures 7 and 8     by a SSL that crossed the Amazon basin in January of 2005 producing a large number of windthrows in the Manaus region [3]. However, only four NSLs [17] were recorded that month, the smallest number recorded during any month in 2005 [53]. We identified several SSLs (November and December 1998, January and February 2002, January 2004 and January 2005) using the TRMM3h data ( Figure 6). Interestingly, these years also contain a high number of windthrows, particularly HY1998-1999, a La Niña year. A peak in windthrow frequency was also observed during HY1999-2000, another La Niña year. The synoptic features associated with the November 1998 SSL event are shown in Figures 7 and 8 [44] were used to plot the displayed meteorological fields. Figure 7a shows a convergence zone extending from the Southwestern Amazon coupled with a cold front that extends to Southeastern Brazil as well as an easterly flow over the equatorial Atlantic Ocean associated with the subtropical high. This easterly flow is deflected by the Andes Mountain resulting in a northerly low level jet (due to the potential vorticity conservation) toward the convergence zone, creating a moisture channel linking this flow back to the southern equatorial Atlantic Ocean, the southeastern Amazon basin and southeastern Brazil. All of these features characterize the typical synoptic-scale pattern associated with a cold front that reaches southeastern Brazil during the austral summer and yields an organized northwest-southeast oriented band of deep convection in southwestern Amazonia. When these features remain stationary for at least four days, they become the South Atlantic Convergence Zone (SACZ) [54,55]. However, the SACZ can also be triggered by the subtropical jet stream (~at 30° S, Figure 7b) which occurs as a consequence of the enhancement of the upper troposphere anti-cyclonic circulation over the Amazon and Central Brazil [56].
An important dynamical feature responsible for the severity, organization and longevity of squall lines is a deep and moderate-to-strong low-level vertical shear that occurs perpendicular to the squall line and controls the ambient flow in which the squall systems are embedded. This essential feature is present in all SSL cases analyzed (represented in Figure 8a). The low-level shear of the wind is oriented northeastward. Thus, the northeastward propagation of the SSL systems can be explained by the lower-troposphere wind shear displayed in Figure 8a. The uniformity of the lower-troposphere vertical shear along the squall line region displayed in Figure 8a  Figures 7a and 8b also show that the SW-NE directed low-level vertical wind shear appears to be a result of the large-scale wind regime over the Amazon, which is associated with a cold front that extends to southeast Brazil. This regime is characterized by northwesterly winds in the lower troposphere, which transport significant moisture toward the NW-SE oriented convergence zone, and a westerly regime in the mid-troposphere (Figure 8b). It is also evident that the mid-troposphere The National Centers for Environmental Prediction (NCEP) Reanalysis data [44] were used to plot the displayed meteorological fields. Figure 7a shows a convergence zone extending from the Southwestern Amazon coupled with a cold front that extends to Southeastern Brazil as well as an easterly flow over the equatorial Atlantic Ocean associated with the subtropical high. This easterly flow is deflected by the Andes Mountain resulting in a northerly low level jet (due to the potential vorticity conservation) toward the convergence zone, creating a moisture channel linking this flow back to the southern equatorial Atlantic Ocean, the southeastern Amazon basin and southeastern Brazil. All of these features characterize the typical synoptic-scale pattern associated with a cold front that reaches southeastern Brazil during the austral summer and yields an organized northwest-southeast oriented band of deep convection in southwestern Amazonia. When these features remain stationary for at least four days, they become the South Atlantic Convergence Zone (SACZ) [54,55]. However, the SACZ can also be triggered by the subtropical jet stream (~at 30 • S, Figure 7b) which occurs as a consequence of the enhancement of the upper troposphere anti-cyclonic circulation over the Amazon and Central Brazil [56].
An important dynamical feature responsible for the severity, organization and longevity of squall lines is a deep and moderate-to-strong low-level vertical shear that occurs perpendicular to the squall line and controls the ambient flow in which the squall systems are embedded. This essential feature is present in all SSL cases analyzed (represented in Figure 8a). The low-level shear of the wind is oriented northeastward. Thus, the northeastward propagation of the SSL systems can be explained by the lower-troposphere wind shear displayed in Figure 8a. The uniformity of the lower-troposphere vertical shear along the squall line region displayed in Figure 8a  Figures 7a and 8b also show that the SW-NE directed low-level vertical wind shear appears to be a result of the large-scale wind regime over the Amazon, which is associated with a cold front that extends to southeast Brazil. This regime is characterized by northwesterly winds in the lower troposphere, which transport significant moisture toward the NW-SE oriented convergence zone, and a westerly regime in the mid-troposphere (Figure 8b). It is also evident that the mid-troposphere westerly wind regime over the Amazon region is associated with the mid-troposphere cyclonic circulation, which provides dynamical support for development of the surface cyclonic circulation associated with the cold front (Figure 8b). In cases when the mid-troposphere westerly flow over the South Amazon is weak, the shear is directed northward and the propagation of the SSL is also nearly northward (figures not shown).

Discussion
In general, we found that the rainy season (SONDJF) has a higher occurrence of windthrows than the dry season (MAMJJA), a pattern concurrent with extreme convection (highest in OND, lowest in AMJ) in Central Amazonia [57]. Seasonal rainfall is regulated by the South American Monsoon System (SAMS) [58][59][60]. The southern hemisphere sector of tropical South America exhibits strong seasonal variation in precipitation and large-scale circulation even though such circulation patterns are not totally reversed as is typical of Northern Hemisphere monsoon systems. The circulation pattern of the SAMS is not fully reversed because the zonal asymmetries of the dry-season circulation are rather weak and, consequently, the mean winter circulation is strongly dominated by the zonally symmetric component of the general circulation. Therefore, although the zonal asymmetry of the large-scale circulation during the dry season is rather weak and thus overwhelmed by the zonally symmetric Hadley circulation, the strong zonal asymmetries characterizing the large-scale circulation during the wet season make it possible to extract a canonical pattern of seasonal variation of the tropical South American large-scale flow that resembles a typical monsoon system. da Silva and Carvalho [61] have defined an Empirical Orthogonal Functions (EOF) based index to characterize the mechanism associated with the intraseasonal and interannual variability of the SAMS. The SAMS has been a topic of intense research throughout the last decade, exhibiting significant inter-annual, intra-seasonal, synoptic and diurnal variability [58,62].
The SAMS has a strong seasonal signal which produces large amounts of rainfall in the austral summer associated with the South Atlantic Convergence Zone (SACZ) [58,59,[63][64][65]. However, apart from the seasonal cycle, studies have shown that the SAMS also exhibit a strong intraseasonal modulation during the austral summer, which is characterized by two distinct phases: an active SACZ and an inactive SACZ [59,63]. During the active phase of SACZ, the diurnal cycle is weak and, although precipitation is abundant, it is mostly stratiform. In contrast, during the inactive phase of the SACZ, the diurnal cycle is stronger and precipitation occurs in the form of deep convection [66,67]. Therefore, windthrows appear to be associated with convective rainfall prevailing in the inactive phase of the SACZ during the austral summer. Furthermore, this convective precipitation associated with windthrow occurrence might be due to organized deep convection activity related to multicell storms that are part of squall lines, since ordinary or single cell storms are unlikely to yield strong gusty winds at the surface [68,69]. In particular, as previously discussed, squall systems embedded in environments with moderate or strong vertical shear in the lower troposphere are the mostly likely candidates to yield the observed windthrows, since the vertical shear in the lower troposphere enhances convective cells organization and, consequently, the intensity of downdrafts, making downbursts more likely [68,69]. This finding is supported by previous studies that examine windthrow formation in Amazonia [1,3].
Regarding squall lines, NSLs are much more common in Amazonia than SSLs and have therefore been more extensively studied. However, SSLs are more frequent than their previously reported 40-50 year return period [23]. Our results suggest that SSLs that originate in southwestern Amazonia and propagate northeastward across the basin are more likely to occur during the wet season ( Figure 6). Furthermore, these SSLs exhibit features typical of mid-latitude squall lines, which are usually generated in the front's warm sector hundreds of kilometers ahead of the surface cold front (Figures 7  and 8). In this context, these SSLs appear to be associated with a warm conveyor belt with forward lift that develops parallel to the cold front along the vanguard and propagates perpendicular to the cold front ahead of it. Therefore, one possible instability mechanism triggering SSLs may be the anomalous moisture convergence and conditional instability that results in deep convection formation in a band-like structure in the front's vanguard. This instability mechanism is similar to that responsible for triggering pre-frontal squall lines in the mid-latitudes [70][71][72][73] and requires both moisture convergence and a surface temperature gradient related to a frontogenic process [74]. The linear forcing mechanism that triggers these squall systems appears to be associated with a conditional symmetric instability related to anomalous frontogenesis development at latitudes around 15 • S and 10 • S. The synoptic-scale flow regime associated with the propagation of the cold front toward the equator yields a directional lower troposphere vertical wind shear. This shear occurs perpendicular to the squall lines and is responsible for the propagation, organization, and longevity of these SSLs. Furthermore, like NSLs, SSLs can also interact with diurnally varying forcings along their trajectories over the Amazon basin, such as river breeze circulations. This interaction may enhance their amplitude at specific locations and times. A full investigation of all mechanisms triggering SSL formation, longevity, propagation and intensity is beyond the scope of this study.
We found that a large number of windthrows were observed even during years with low rainfall. This suggests that a higher temporal resolution analysis is needed to capture the types of systems producing intense rainfall and windthrows (Supplementary Materials Figure S1). Higher temporal and spatial rainfall data is needed. The TRMM3h data appear to be insufficient given that a system can cross the entire Landsat scene within the 3 h TRMM3h window. Furthermore, the spatial resolution of TRMM3h data (0.25 • ) is orders of magnitude larger than most frequent windthrows (range of 1 tree to 100 ha). The current Global Precipitation Measurement data (GPM) has a 30 min time step and 0.1 • resolution [75], which may help to provide more detailed analysis. However, even with the proper rainfall data, a constraint will remain due to the dating uncertainty of windthrows (Section 2.4) in long time series' across the entire Amazon.
ENSO [58,63,76,77] affects the interannual variability of rainfall in the Amazon. During the El Niño (La Niña) phase of ENSO, there is a higher (lower) than average sea surface temperature in the equatorial Pacific ocean [26] which is associated with lower (higher) than average rainfall over Central Amazonia [26,78]. The decrease (increase) of rainfall during the El Niño (La Niña) is associated with the weakening (strengthening) of easterly circulation in the lower troposphere resulting in a decrease (increase) of humidity transport toward the Amazon, decreasing (increasing) the potential for convection. Though during some La Niña years we found a higher number of windthrows ( Figure 5), we did not find an association between ENSO and windthrows. This may be related to the fact that the occurrence of extreme convective system does not follow the ENSO variability ( Figure S1). We emphasize that, despite the increased convection during La Niña years, a longer time series is needed to establish an association between windthrows and ENSO. Though Landsat imagery are available since the 1970s, the TRMM3h data is only available since 1998 and the GPM data since 2014. TRMM3h together with the GPM and Landsat provide at best about 20 years of data, which may still not be enough for a rigorous statistical analysis of the association between windthrows and ENSO.
Long-lived northerly squall lines are associated with low level jets (LLJs), a phenomenon expected to be more frequent under a warming climate [79]. LLJs may also trigger SSLs, given that they are essential synoptic-scale features that propagate cold fronts toward Southeast Brazil. To our knowledge, current ESMs are unable to represent windthrows. However, changes in the frequency of windthrows are important within the context of forest shifts that will affect the terrestrial carbon budget, and, therefore, feedback with climate [14,80]. Thus, predicting the occurrence of the meteorological systems that produce windthrows holds important ecological consequences and may prevent fatalities [81] and economic losses [82] like those that occurred following the January 2005 SSL event [3].
It should be emphasized that the interaction of wind and trees is nonlinear and involves complex processes and the integration of disciplines such as soil science, physiology, ecology, biomechanics, and meteorology [6,11,83,84]. Wind, wind loading, gusts of winds, tree stature, tree species (and associated characteristics such as tree crown shape and density, root architecture and shape, and wood density), topography, soil, cumulative processes, etc. all play a role in the production of windthrows [11,[83][84][85][86][87][88] resulting in tree failure at a lower wind speed than expected [1,3]. Amazonia covers about 5.3 million km 2 , and has many environmental and functional gradients [89,90]. For instance, western Amazonia maintains higher rainfall and wood productivity compared with the long dry-season and less productive eastern Amazonia [89,91]. The environmental variability across the basin made it necessary to divide it into regions (e.g., [10,92,93]). The focus of our study is Central Amazonia because this area represents a large fraction of Amazonia. Identifying windthrows allowed us to discover a higher frequency of SSLs across the Amazon than previously reported. These squall lines have important implications not only for meteorology but also for disciplines like plant physiology and ecology. Our study provides a method that can easily be replicated in other areas of the Amazon to later integrate a basin wide perspective.
Some important aspects of the variability of windthrows remain unexplored due to limitations associated with image availability, resolution, and focus of this study. For instance, the monthly variability of windthrows will be an important factor in more closely linking windthrow occurrence and rainfall, especially due to the aforementioned intraseasonal modulation of the SAMS. A case by case study is needed to determine the atmospheric characteristics that cause downbursts and windthrows in Amazonia. The use of Landsat imagery limited the size of windthrows studied, and, therefore, the whole gradient of windthrow variability-from smaller windthrows (<5 ha) to more frequent windthrows (1 single windthrown trees)-remains to be studied. Furthermore, the tropical North Atlantic SST has a stronger influence over southern Amazonian rainfall during the dry season and when ENSO has limited activity (as in 2005 [39]), and the South Atlantic SST has limited influence on rainfall over the Atlantic coast of South America and the southern edge of the basin during the early dry season [94,95]. Thus, Atlantic SSTs mainly influence southern Amazonian rainfall where SSLs are formed, and, therefore, an association between Atlantic SSTs and windthrows should not be ruled out. However, to our knowledge, there are no studies addressing this association.

Conclusions
Our study shows that, although windthrows occur all year long, they have a seasonal and interannual variability driven by severe convective systems. The variability of windthrows is somewhat mirrored in the variability of annual rainfall. In general, a higher number of windthrows occurs during the rainy season (SONDJF). However, we did not find an association between windthrows and ENSO over the study period. Severe convective events associated with mesoscale convective systems, such as squall lines, drive the observed variability. We found that southerly squall lines have a higher frequency of occurrence and a greater effect on windthrows than previously reported. These systems deserve special attention given the large amount of rainfall they produce and their resulting ecological impacts. Because windthrows are an important driver of forest structure and dynamics, our study emphasizes the need to include windthrows in ESMs in order to reduce the uncertainties of climate predictions.