Assessing the relative importance of dry-season incoming solar radiation and water storage dynamics during the 2005, 2010 and 2015 southern Amazon droughts: not all droughts are created equal

Three severe droughts impacted the Amazon in 2005, 2010, and 2015, leading to widespread above-average land surface temperature (LST) (i.e. positive thermal anomalies) over the southern Amazon in the dry season (Aug–Sep) of these years. Below-average dry-season incoming solar radiation (SW↓) and terrestrial water storage anomaly (TWSA) were simultaneously observed in 2005 and 2010, whereas the opposite was observed in 2015. We found that anomalies in precipitation (P), SW↓, and TWSA combined can well explain dry-season thermal anomalies during these droughts (average R2–0.51). We investigated the causes for opposing anomalies in dry-season SW↓ and TWSA, and found different hydro-climatological conditions preceding the drought-year dry seasons. In 2005 and 2010, P was considerably below average during the wet-to-dry transition season (May–July), causing below-average TWSA in the dry season that was favorable for fires. Increased atmospheric aerosols resulting from fires reduced solar radiation reaching the ground. In 2015, although below-average dry-season P was observed, it was above the average during the wet-to-dry transition season, leading to reduced fires and aerosols, and increased dry-season SW↓. To further examine the impact of opposite hydro-climatological processes on the drought severity, we compared dry-season LST during droughts with the maximum LST during non-drought years (i.e. LSTmax) for all grid cells, and a similar analysis was conducted for TWSA with the minimum TWSA (i.e. TWSAmin). Accordingly, the regions that suffered from concurrent thermal and water stress (i.e. LST > LSTmax and TWSA < TWSAmin) were identified. These regions are mainly observed over the southeast in 2005 and southern Amazon in 2010. In 2015, large-scale dry-season thermal stress was found over central and southeast Amazon with little water stress. This study underlines the complex interactions of different hydro-climatological components and the importance of understanding the evolution of droughts to better predict their possible impacts on the Amazon rainforest.


Introduction
The Amazon is the most extensive tropical rainforest on Earth, playing an essential role in water, carbon cycle, ecosystems, and climate (Cox et al 2008, Phillips et al 2009, Davidson et al 2012, Gatti et al 2021).The Amazon has been confronted with extreme events caused by climate change and anthropogenic activities, i.e. agricultural expansion and deforestation (Davidson et al 2012, Silva Junior et al 2020).These threats lead to forest degradation and losses (Lewis 2006), which could further intensify carbon emissions and climate change (Phillips et al 2009, Harris et al 2021).The resilience of the Amazon rainforests to these stresses is of widespread concern (Boulton et al 2022).
Three mega-droughts hit this region in just over a decade, including 2005 (Saleska et al 2007), 2010 (Lewis et al 2011, Marengo et al 2011) and 2015(Jiménez-Muñoz et al 2016, Aragão et al 2018).Extensive studies have sought to determine how climate mechanisms induced these anomalous droughts (Zeng et al 2008, Jiménez-Muñoz et al 2016, Marengo and Espinoza 2016).These severe droughts led to reductions in photosynthesis (Yang et al 2018), greenness and growth rate (Bonal et al 2016, Liu et al 2018, Xu et al 2011, Zhao et al 2017), and increases in tree mortality and flammability (Phillips et al 2010, Feldpausch et al 2016, Zhang and Liu 2022).The Amazon rainforest has been losing its resilience, particularly in regions receiving less precipitation that are closer to human activities (Boulton et al 2022).Investigations on the anomalies and dynamics of the hydro-climatological conditions of past droughts are crucial to enhance our understanding of the hydro-climatological extremes impact on drought-stricken areas and improve the robustness of the forecasting models of rainforest under climate change.
The Amazonia suffered thermal and water stress during the droughts.Many studies have investigated hydro-climatological factors of the three droughts using field observed and/or remotely sensed data.Warming can enhance droughts from precipitation deficit (van der Schrier et al 2013) and increased potential evapotranspiration (Jiménez-Muñoz et al 2013).Land surface temperature (LST) can characterize energy fluxes and water budgets (Mannstein 1987, Park et al 2005), which are commonly used for describing thermal conditions.Anomalously high LST were observed in the dry season (August-September) in all three droughts (Liu et al 2022).Concurrently, the southern Amazon became drier in three droughts based on river discharge levels and precipitation records (Marengo et al 2008, Coelho et al 2012).Terrestrial water storage anomaly (TWSA) declined from May to October in southern Amazonia in 2005 and 2010 droughts and was below the longterm average of 2003-2010 (Zhang and Liu 2022), while TWSA in 2015 were higher than the long-term average (Yang et al 2018).Incoming surface solar radiation (SW↓) is also one of the key drivers for LST dynamics (Hulley et al 2019).It controls tree growth and canopy greenness in rainforests (Guan et al 2015).The amount of SW↓ was less in the dry season of 2005 and 2010 than the long-term dry-season average (Aragão et al 2007, Anderson et al 2010, Samanta et al 2010, Zhao et al 2017, Zhang and Liu 2022), while SW↓ was higher in 2015 (Yang et al 2018).
However, the causes of dry-season thermal stress and the contribution of the relevant hydroclimatological factors have not yet been fully explored.Moreover, the causation of the opposite anomalies in dry-season SW↓ and TWSA among these three droughts requires further investigation.Besides the anomalies of hydro-climatological factors, the severity of thermal stress and water shortage are of high importance.Identifying regions that were simultaneously affected by both thermal stress and water stress is needed, as is quantifying the level of each stress.The response of the rainforest to severe thermal and water stresses over these regions are key elements in climate models, thus identifying these severely affected areas is fundamental.
Accordingly, the overarching objectives of this study are three-fold.The first objective is to quantify the contribution of hydro-climatological factors driving the dry-season thermal anomalies.The second is to explain the causes of the opposite pattern of anomalies in dry-season SW↓ and TWSA of these three droughts.The third objective is to quantify the severity of the dry-season thermal and water stresses in the drought years and identify their affected regions.Achieving these objectives will broaden our knowledge of how the rainforests respond to climate perturbations, and the management and protection of rainforests on Earth.

Satellite-derived datasets
We focused on the southern Amazon rainforest (0 1), which experienced all three droughts in 2005, 2010 and 2015 (Marengo et al 2008, Xu et al 2011, Liu et al 2018); the northern Amazon was less impacted by the 2005 drought (Marengo et al 2008).Several sources of satellite and re-analysis data were used to understand the characteristics and drivers of the thermal and water storage anomalies (see table 1), including: (i) LST from the daytime overpass (13:30 pm) of the Moderate Resolution Imaging Spectroradiometer (MODIS) onboard Aqua MYD11C3 product (Collection 6.1) (Wan 2014), which in areas of high tree cover (%) All hydro-climatological data cover from January 2003 through to December 2015 (and to December 2016 in the supplementary information S1 and S2).They were resampled to achieve common 0.1 • spatial resolution, with either bilinear interpolating or aggregating their initial resolutions to enable a direct comparison, except active fires.MCD14ML is a monthly fire location dataset containing the date of occurrence, geographic location (latitude, longitude), and detection confidence.Active fires with high detection confidence (>80%) were retained to reduce false positive (Giglio et al 2020).All active fires identified in the same grid cell (0.1 • ) of the same day were added up, and then the total number of active fires per month for each 0.1 • grid cell were accumulated.

Contribution of dry-season radiation and water storage anomalies to thermal anomalies during the three Amazon droughts
Herein, August and September are considered the dry season of the southern Amazon.The regional average monthly LST in August and September is higher than July based on long-term timeseries of LST observed in the studied region (figure S1).The selection of August to September, rather than the typical July to September Amazon dry-season, aimed to better understand the extreme thermal events during droughts.First, we derived the dry-season anomalies in LST and associated hydro-climatological variables.The anomaly is the departure of the dry-season value from the averaged value in the dry seasons during the reference period excluding the drought-years (i.e.all non-drought years from 2003 through 2015, excluding 2005, 2010 and 2015).
For instance, the LST anomaly is calculated as: where LST ano refers to the per-pixel anomaly in observed LST in the dry season (August and September) of a particular drought year, from which the averaged dry-season LST (LST) grid of all nondrought years is subtracted.Additionally, we applied this method to precipitation (P), incoming surface solar radiation (SW↓) and TWSA to derive their anomalies in the dry season from 2003 through 2015 for each grid cell over the southern Amazon.It is noted that we also tested using hydro-climatological variables for the period of 2003-2016 via the same method (supplementary information S1 and S2).
Second, we simulated dry-season LST anomaly during 2003-2015 for each grid cell using other hydro-climatological variables and linear The regression model was fit using the least squares method.Coefficient of determination (R 2 ) and root mean square error (RMSE) were calculated to assess the simulation performances.In addition, the linear regression was also conducted for the period of 2003-2016 (supplementary information S1 and S2).
Furthermore, we compared the observed and simulated dry-season LST anomaly in drought years (i.e. 2005, 2010 and 2015) for each grid cell.The R 2 was used to evaluate their spatial consistency.We also quantified the individual contribution of anomalies in P, SW↓ and TWSA to the simulated LST anomaly, i.e. a × P ano , b × SW↓ ano , and c × TWSA ano , respectively, which enables us to better understand how the influence of these variables changes between the three droughts.

Causes for opposite dry-season radiation and water storage anomalies during the three Amazon droughts
To identify the possible causes for opposite dryseason radiation and water storage anomalies (see table 1), we examined the hydro-climatological conditions in the preceding wet-to-dry transition season (May to July) during three droughts, including P, SW↓, TWSA, AOD, COT and active fires.We compared their values and anomalies from May through September in 2005, 2010 and 2015, respectively, to better understand the processes resulting in the opposite anomalies.

Severity of dry-season thermal and water stress during the three Amazon droughts
Herein we examined whether LST and TWSA values in drought years exceeded their ranges during the non-drought years (see figure S3 for example).Following the method illustrated in figure S3, we first calculated LST max and TWSA min for all grid cells.Then for each grid cell, we compared the dry-season LST in drought years with LST max to determine the regions with LST > LST max and the magnitudes of exceedance.A similar analysis was conducted between dry-season TWSA and TWSA min .Accordingly, the regions with concurrent thermal and water stress (i.e.LST > LST max and TWSA < TWSA min ) during the three Amazon droughts were identified.Figure 3 compares the performance of linear regression models to simulate the dry-season LST anomaly during 2003-2015 over each grid cell, using: (i) only P ano ; and (ii) the combination of P ano , SW↓ ano and TWSA ano , respectively.The average R 2 and RMSE between the observed and simulated LST ano using linear regression of only P ano are 0.27 and 0.61 K, respectively.When using the combination of P ano , SW↓ ano and TWSA ano , the average R 2 value increases to 0.51 and RMSE value decreases to 0.47 K. Overall, the combination of three hydrometeorological variables outperforms the models using single variable only (i.e.(i) P ano ; (ii) SW↓ ano ; (iii) TWSA ano , as provided in the figure S4) and the combinations of two variables (i.e.(i) P ano and SW↓ ano ; (ii) P ano and TWSA ano ; or (iii) SW↓ ano and TWSA ano , as provided in the figure S5).The spatial distributions of coefficients of P ano (a), SW↓ ano (b), TWSA ano (c) and the intercept (d) obtained from the trivariate linear regression are provided in figure S6.Across most regions, the coefficients a and c are negative, while b and d are predominantly positive.Positive coefficients mean that the variable positively contributes to a higher LST ano , and vice versa.

Contribution of dry-season radiation and water storage anomalies to thermal anomalies during the three Amazon droughts
Next, we focus on the simulated dry-season LST ano during the drought years via the linear regression model which combines P ano , SW↓ ano and TWSA ano .The simulated LST ano reasonably captured the spatial patterns of observed LST ano in the drought years (figures 4(a)-(c)).The R 2 of simulated LST ano and observed LST ano are 0.53, 0.49 and 0.61, for the 2005, 2010 and 2015 droughts, respectively (figures 4(d)-(f)).We assessed the collinearity of three variables by examining the variance inflation factors (VIF), revealing low values (less than 4) for individual variable across most grids (figure S7).Based on the VIF threshold for collinearity (less than 5) (James et al 2013), the collinearity of variables (P ano , SW↓ ano , TWSA ano ) can be considered low in the grid-based regression.We also identified the contributions of P ano , SW↓ ano , and TWSA ano to the simulated dry-season LST ano in the drought years (figures 4(g)-(i)).The contribution of dryseason P ano to positive LST ano is comparable during three drought years.The contributions of SW↓ ano and TWSA ano show different responses across the three droughts.SW↓ ano had an opposing simulation contribution between the first two droughts and the third one: that is, negative mean SW↓ ano contribution for 2005 and 2010 droughts (figures 4(g) and (h)), while positive mean SW↓ ano contribution for 2015 (figure 4(i)).The TWSA anomaly also yielded an opposing characteristic across the three droughts, as TWSA anomaly positively contributed to LST anomaly simulation in 2005 (figure 4(g)) and 2010 (figure 4(h)), while negative in 2015 (figure 4(i)).Note SW↓ ano and TWSA ano compensated each other during all three droughts; that is, when one exhibited a positive contribution the other had a negative contribution and vice versa.In sum, the contributions of anomalies in precipitation, solar radiation and water

Causes for opposing dry-season radiation and water storage anomalies during the three Amazon droughts
In section 3.1, we observed an opposing contribution of the dry-season SW↓ and TWSA anomalies across the three droughts.To understand this phenomenon, we investigated the anomalies of relevant hydro-climatological variables during the dry season (August and September) and the antecedent wet-todry transition season (May to July) of three droughts.Six variables were examined, namely P, TWSA, SW↓, AOD, COT and active fires (see table 1).Monthly regional averages of these six variables and their anomalies from May to September are compared in figure 5.
Regional average precipitation decreases from May to August in droughts and non-drought years alike (figure 5 Incoming surface solar radiation (SW↓) (figures 5(e) and (f)) normally continues to increase from May to September.During May to July, SW↓ in the droughts years was very similar to the non-drought years.However, in 2015 August and September SW↓ was much higher than average, while To understand the possible causes of these SW↓ patterns in the drought years, the dynamics of aerosols and clouds were examined using AOD and COT, respectively (figures 5(g)-( j Spatially, the active fires mainly occurred in the southeastern region where the tree cover density is relatively low (figure 1).Silveira et al (2022) highlighted that the anomalies of fire extent were associated with deforestation rather than water deficit anomalies.
The fires-induced aerosols, together with cloud particles, influence SW↓ to the canopy (Bian et al 2021).Figure 6   (i) are the same as (g), but for August-September 2010 and 2015, respectively.P represents precipitation, SW↓ represents solar radiation, and TWSA represents terrestrial water storage anomaly.For the box-whiskers, the mid horizontal is the median, the upper and lower horizontal lines (defining the box) are the 75% and 25%, with the whiskers being the maximum and minimum.2022).However dry-season AOD was below-average in 2015 due to fewer fires and below-average cloudiness.This may be linked to precipitation occurrence in the previous months, including the wet-to-dry transition season, resulting in adequate water storage in the landscape.With wetter antecedent conditions, the rainforest was less flammable (Nolan et al 2020), even though the thermal stress was similar to 2005 and 2010 droughts.

Severity of dry-season thermal and water stress during the three Amazon droughts
To quantify the severity of the thermal and water stress during the dry season of these three drought years, we first calculated LST max and TWSA min for all grid cells in the southern Amazon (figures 7(a) and (b)).As illustrated in figure S3, LST max represents the maximum average monthly LST in the non-drought years, while TWSA min represents the minimum average monthly TWSA in non-drought years.
In 2005, the dry-season LST was higher than LST max (figure 7(a)) over one thirds of the study area (figure 7(c)), while TWSA was lower than TWSA min (figure 7(b)) over 60% of this area (figure 7(d)).This resulted in simultaneous thermal and water stress (using the metrics of 'stress' defined herein) being experienced by over 30% of southern Amazon, primarily over the western part of the region (see hatched area in figures 7(c) and (d)).In 2010, the region that suffered from both thermal and water stress became broader ∼50% (figures 7(e) and (f)), although the magnitude of water stress was less severe than 2005 (figure 7(f)).In 2015, dry-season LST was higher than LST max over the eastern 60% of the study area, but TWSA was higher than TWSA min over all grid cells (figures 7(g) and (h)) indicating that no   (2005, 2010, and 2015).Mean wind direction and speed at 850 hPa are shown in (g)-(i).It is noted that the active fires occurred in both high-density (⩾60% tree cover) and low-density rainforest (<60% tree cover) in the southern Amazon.
water stress (as defined here) was experienced in the 2015 drought.The identification of areas experiencing concurrent extreme LST and TWSA are important for understanding the response of rainforests under extreme climate.
We noticed that the 2015/2016 drought duration was from August 2015 to July 2016 (Yang et al 2018).
In this study, we focused on the dry-season (August-September) of droughts.We also conducted the sensitivity analysis by using the timeseries of 2003-2016 (see section S1, figures S8-13).During the 2016 dryseason, the studied region encountered above-average LST, with a concentration of higher values in the western part.Dry-season LST ano was comparatively lower For each of the three droughts independently, areas with concurrent positive LST minus LSTmax and negative TWSA minus TWSA min were hatched.In the bottom-right of each subplot on the left column, the first number (i.e. to the left of forward slash) is the percentage of southern Amazon high-density rainforest that experienced thermal stress (i.e.LST minus LSTmax is positive), and the second number (i.e. to the right of forward slash) is the percentage of southern Amazon high-density rainforest that experienced both thermal and water stress.The number in the right column are similar to left column, except that the first number is the percentage of southern Amazon high-density rainforest that experienced water stress (i.e.TWSA minus TWSA min is negative).The legend below (g) also applies to (c) and (e), with the legend below (h) also applying to (d) and (f).uncertainties resulting from precipitation, we repeated the analyses using a different precipitation dataset, Climate Hazards group Infrared Precipitation with Stations (CHIRPS) precipitation (Funk et al 2015).CHIRPS precipitation data combines estimations from rain gauge and satellite observations.The results generated from using CHIRPS (see section S2, figures S14-17) concord the analyses using IMERG, which enhances the reliability of our results.
With climate change, the magnitude, extent and duration of extreme events may be amplified, particularly in the vulnerable rainforest regions (Marengo et al 2018).If future El Niño frequency increase, enhanced extreme hot droughts are forecasted to be more extensive in Amazonia (Panisset et al 2018, Yang et al 2018), with reductions in precipitation (Hilker et al 2014, Parsons 2020), and doubled risks of extreme fire-weather conditions by the end of 21st century (Touma et al 2021).Under such future circumstances, the concurrent thermal and water stress are expected to severely affect a wider range of regions in the Amazon catchment.Limiting deforestation and active fires can intervene rainforest losses in the unabated climate warming (Malhi et al 2009).
Worldwide, droughts have significant impacts on rainforests in tropical regions (e.g.Amazonia, Congo, southeastern Asia) (Bastos et al 2023).Our analysis of antecedent hydro-climatological conditions preceding to dry-season and the quantification of concurrent thermal and water stress can be generalized to understand what may happen to rainforests globally.The past severe droughts in these vulnerable regions may provide a proxy for future climate conditions, and the response and sensitivity of rainforest during these droughts may be reproducible in 21st century climates.

Conclusion
This study focused on understanding the causes of dry-season positive anomalies in LST in the southern Amazon rainforest during the 2005, 2010, and 2015 droughts.The study used satellite remote sensing and trivariate linear regression to explore the spatial and temporal dynamics of hydro-climatological drivers and their influences in LST anomalies.The anomalies in precipitation, terrestrial water storage, and incoming solar radiation combined can reasonably explain dry-season LST anomalies.The contributions of anomalies in TWSA and SW↓ were opposite in different years.TWSA anomalies contributed positively to dry-season LST anomalies in 2005 and 2010, but negatively in 2015.The contribution of SW↓ anomalies to dry-season LST anomalies is negative in 2005 and 2010, but positive in 2015.Therefore, the underlying causes of the three droughts are different.
The different hydro-climatological conditions preceding the dry-season were important to explain the onset of opposite patterns in TWSA and SW↓ anomalies between drought years.For instance, the precipitation was above-average and the catchment was in wet conditions in the wet-to-dry transitional season, leading to less fires and aerosols in the dry season, which subsequently resulted in above-average SW↓ in 2015.We can well explain how below-average radiation led to above-average temperature during droughts.
A new threshold-based method was adopted to quantify the severity of droughts, where the dryseason LST and TWSA exceeded their normal ranges.The simultaneous thermal and water stress was mainly observed over the southeast in 2005 and southern Amazon in 2010.The concurrent regions increased by around 19% from 2005 to 2010.In 2015, large-scale dry-season heat stress was observed over central and southeast Amazon with little water stress.
Future Amazon droughts are projected to be more frequent and severe.Our study highlights the importance of understanding the evolutions of extreme hydro-climatological events in the rainforest.It will contribute to making better predictions of future rainforest responses to climate perturbations, bolstering the robustness of climate forecasting models, and elevating the management and protection of rainforests on a global scale.The findings and methodology from this study can be generalized to other rainforest regions to understand and characterize drought occurrence in this ecosystem worldwide.

Figure 1 .
Figure 1.Spatial distribution of tree cover (%) averaged over 2003-2015 in southern Amazonia.Grid cells with 60% and above tree cover based on the vegetation continuous fields in the Collection 6 MOD44B product (Dimiceli et al 2015) define the study area.
Spatial distributions of dry-season anomalies in LST, P, SW↓, and TWSA in 2005, 2010 and 2015 are shown in figure 2. All three droughts experienced widespread positive anomalies in LST and negative anomalies in P (figures 2(a)-(f)), although the regions with the strongest anomalies vary between droughts.Striking discrepancies are observed when comparing the SW↓ and TWSA across droughts.Negative anomalies in SW↓ are found over three quarters of southern Amazon in 2005 and 2010, particularly the southernmost region, whereas a positive SW↓ anomaly is seen over 80% of the study area in 2015 (figures 2(g)-(i)).The anomalously decreased SW↓ in 2005 and 2010 and increased SW↓ in 2015 are consistent with previous studies (Samanta et al 2010, Liu et al 2018, Yang et al 2018).While the entire region experienced negative anomalies in TWSA in 2005 and 2010, more than 60% of southern Amazon showed above-average TWSA in 2015 (figures 2(j)-(l)).The opposing dryseason TWSA anomalies during the three droughts are supported by the river flow measurements: these were low in the 2005 and 2010 dry seasons (Marengo and Espinoza 2016), while being above-average in the 2015 dry season (Yang et al 2018).
(a)).However, precipitation in 2005 and 2010 was constantly lower than the average of non-drought years consecutively for these five months (figure 5(b)).Although negative P ano were also observed in August and September of 2015 (with the August 2015 P ano being near-zero), the study region mainly experienced above-average precipitation (positive anomalies) during the wet-to-dry transition season (May to July) in 2015.Therefore, the 2005 and 2010 antecedent conditions of precipitation differed markedly from those experienced in 2015.As a result, TWSA in 2015 was constantly higher than the non-drought years, while the other two droughts had below-average TWSA (figure 5(c)).TWSA anomalies were above 50 mm from May to September 2015, but below −70 mm each month in 2005 and 2010 (figure 5(d)), suggesting that the southern Amazonia had adequate water storage before the dry season in 2015, while water shortage already commenced in the wet-to-dry transition season for both 2005 and 2010.The prolonged water stress started from wet-todry transition season affected soil moisture availability and caused tree mortality and rainforest dieback (Wey et al 2022).In contrast, adequate wet-to-dry water storage in 2015 provided sufficient soil moisture for the rainforest to survive in that dry season (Meng et al 2022).

Figure 3 .
Figure 3. Relationship between observed and simulated dry-season LST anomalies during 2003-2015 (left column).Spatial distribution of (a) coefficient of determination (R 2 ) and (b) root mean square error (RMSE) between satellite-based LSTano and LST ano−sim from the per-pixel linear regression model with only Pano.(c) Area (%, left y-axis, in black) and average RMSE with standard deviation (K, right y-axis, in blue) for different groups of R 2 values.(d-f) in right column are the same as the left column but use the combination of Pano, SW↓ano and TWSAano in the per-pixel linear regression model.
)). COT in May to July of three drought years fluctuated around the mean of non-drought years in a roughly similar manner (figure 5(g)).In August and September, COT was at (August) or above average (September) in 2005, but below average in 2010 and 2015 (figure 5(h)).For aerosols, AOD was similarly low during May to July in both drought and non-drought years (figures 5(i)-(j)).It dramatically increased in August and September 2005 and 2010 to levels twice the average (figure 5(i)), while AOD was slightly lower than the average in August and September 2015 (figure 5(j)).Furthermore, we found that temporal patterns of AOD anomaly are closely associated with the active fires (figure 5).The number of active fires in August and September of 2005 and 2010 was nearly twice of the average in non-drought years (figure 5(k)), similar to the magnitude of AOD anomaly (figure 5(i)).The extreme high dry-season aerosol concentration in 2005 and 2010 are consistent with previous findings (Bevan et al 2009, Reddington et al 2019) due to large-scale fires.The preceding monthly TWSA (i.e.1-5 months) are significantly associated with the dryseason active fires (Chen et al 2013).In 2015, the number of active fires were slightly lower than the average (figure 5(l)), similar to the AOD (figure 5(j)).
shows the spatial distribution of dryseason anomalies in AOD and COT of the three drought years.In 2005, both AOD and COT were above average (figures 6(a) and (d)), which strongly reduced the radiation reaching the ground and caused below-average SW↓.In 2015, both AOD and COT were below average (figures 6(c) and (f)), which are favourable for radiation reaching the ground and likely caused above-average SW↓.For 2010, AOD was above-average yet COT was below-average (figures 6(b) and (e)); this combination very likely caused the SW↓ anomaly to fall between that of 2005 and 2015.The aerosols associated with biomass burning, were dispersed across the Amazon by the wind (figures 6(g)-(i)), resulting in widespread above-average AOD in 2005 and 2010 that concords with other studies (Martins et al 2018, Zhang and Liu

Figure 4 .
Figure 4. Characterisation of simulated LST anomalies during the three drought years.Parts (a)-(c), respectively, show the spatial distribution of simulated dry-season LST anomalies of (a) 2005, (b) 2010, and (c) 2015 based on the trivariate linear regression model using the combination of Pano, SW↓ano and TWSAano.Scatterplot of observed and simulated dry-season LST anomalies of the three droughts are shown in (d)-(f), being (d) 2005, (e) 2010, and (f) 2015.N represents the number of grid cells used to develop linear regression model.The coefficient of determination (R 2 ) and root mean square error (RMSE) of individual drought years are shown on each sub-part.Box-whisker plot of the simulated LST anomaly is shown in (g), with the contributions of Pano, SW↓ano and TWSAano to simulate LST anomaly in August-September 2005 based on the trivariate linear regression.Parts (h) and(i) are the same as (g), but for August-September 2010 and 2015, respectively.P represents precipitation, SW↓ represents solar radiation, and TWSA represents terrestrial water storage anomaly.For the box-whiskers, the mid horizontal is the median, the upper and lower horizontal lines (defining the box) are the 75% and 25%, with the whiskers being the maximum and minimum.

S
Figure 5. Regional average monthly values and anomalies of (a)-(b) precipitation (P), (c)-(d) terrestrial water storage anomaly (TWSA), (e)-(f) surface incoming solar radiation (SW↓), (g)-(h) cloud optical thickness (COT), (i)-(j) aerosol optical depth (AOD), and (k)-(l) the number of active fires from May to September for the three drought years (2005, 2010, and 2015).The legend on (a) applies to all other line plots, with the legend on (b) applying to all other anomaly bar charts.There is no TWSA data in June 2015 due to satellite battery management issues.This is the only month-year combination for the six variables having missing data.

Figure 6 .
Figure 6.Dry-season anomalies in (a)-(c) aerosol optical depth (AOD), (d)-(f) cloud optical thickness (COT), and (g)-(i) the number of active fires of three drought years(2005, 2010, and 2015).Mean wind direction and speed at 850 hPa are shown in (g)-(i).It is noted that the active fires occurred in both high-density (⩾60% tree cover) and low-density rainforest (<60% tree cover) in the southern Amazon.

Figure 7 .
Figure 7. Extreme values of land surface temperature (LST) and terrestrial water storage anomaly (TWSA) during non-drought years and the exceedance of dry seasons in drought years.Spatial distribution of (a) maximum LST (LSTmax) and (b) minimum TWSA (TWSA min ) during non-drought years.(c) Difference between mean LST in August-September 2005 and LSTmax as shown in (a).(d) Difference between mean TWSA in August-September 2005 and TWSA min as shown in (b).(e)-(f) and (g)-(h) are the same as (c)-(d), but for August-September 2010 and 2015, respectively.It is noted that only positive LST minus LSTmax, and negative TWSA minus TWSA min were shown in (c)-(h).For each of the three droughts independently, areas with concurrent positive LST minus LSTmax and negative TWSA minus TWSA min were hatched.In the bottom-right of each subplot on the left column, the first number (i.e. to the left of forward slash) is the percentage of southern Amazon high-density rainforest that experienced thermal stress (i.e.LST minus LSTmax is positive), and the second number (i.e. to the right of forward slash) is the percentage of southern Amazon high-density rainforest that experienced both thermal and water stress.The number in the right column are similar to left column, except that the first number is the percentage of southern Amazon high-density rainforest that experienced water stress (i.e.TWSA minus TWSA min is negative).The legend below (g) also applies to (c) and (e), with the legend below (h) also applying to (d) and (f).
than during the droughts of 2005, 2010 and 2015.The performance of linear regression for 2003-2016 (R 2 -0.48) were similar to 2003-2015 (R 2 -0.51).The use of various gridded precipitation data can introduce uncertainties over tropical forests(Papastefanou et al 2022).To examine the potentialS Liu et al

Table 1 .
List of satellite and re-analysis datasets used herein.