Estimating vegetation water content from Sentinel-1 C-band SAR data over savanna and grassland ecosystems

Studying vegetation water content (VWC) dynamics is essential for understanding plant growth, water and carbon cycles, and ecosystem stability. However, acquiring field-based VWC estimates, consistently through space and time, is challenging due to time and resource constraints. This study investigates the potential of Sentinel-1 C-band Synthetic Aperture Radar (SAR) data for estimating VWC in natural ecosystems in central Brazil. We assessed (i) how well Sentinel-1 SAR data can capture variations in VWC over three different vegetation types (i.e. dry and waterlogged grasslands, and savannas) and (ii) how the studied vegetation types respond to seasonal dry periods in terms of water content. Field data from 82 plots, distributed across the three vegetation types and revisited in four different seasons, were used to calibrate and validate a model for VWC estimation. The calibrated model, with an R 2 of 0.52 and RMSE of 0.465 kg m−2, was then applied to Sentinel-1 SAR backscatter data to generate monthly VWC maps for grassland and savanna ecosystems at 30 m spatial resolution between April 2015 and September 2023. These maps, combined with rainfall and potential evapotranspiration data, provided insights into how the studied vegetation types respond to water shortage during the dry season at the community scale. More specifically, savannas showed to be better able to retain higher levels of water content during the dry season, probably due to a higher water holding capacity of the woody component together with its deep-root system ability to access deeper groundwater. This research demonstrates the potential of Sentinel-1 SAR data for monitoring VWC in natural ecosystems, allowing for future studies to assess ecosystems’ response to drought events and changes in their functioning, ultimately supporting land management decisions.


Introduction
Vegetation water content (VWC) is a crucial biophysical parameter linked to plant growth, ecosystem dynamics, and to water and carbon cycles (Choat et al 2018, Konings et al 2021).Although some vegetation types are adapted to seasonal dry and wet periods, changes in precipitation patterns due to climate change can affect vegetation stability (MEA 2005, Huang et al 2016, Bernardino et al 2020a, IPCC 2023).In parallel, anthropogenic overexploitation of natural resources can aggravate stability loss (MEA 2005, Barbier et al 2006).Since understanding changes in VWC can contribute to assessing the impacts of climate change and human activities on ecosystem functioning, accurately estimating VWC becomes essential.
Traditional field surveys for VWC data are resource-intensive and limited in space and time, while spaceborne remote sensing provides a valuable alternative (Konings et al 2021), given its widespread coverage and temporal consistency.Several studies were already able to successfully estimate VWC using optical or passive microwave data (e.g.Jackson et al 2004, Grant et al 2012, Bueso et al 2023, Feng et al 2023).However, the Sentinel-1 constellation, in particular, offers valuable information on VWC since they carry Synthetic Aperture Radar (SAR) sensors which are sensitive to the dielectric properties of both vegetation and soil (Vreugdenhil et al 2020, Wang et al 2023).Moreover, the active nature of SAR sensors provides data in high spatial resolution and independent of cloud coverage and sunlight (Hanssen 2001, Konings et al 2021).
Several factors can contribute to the amount of radar backscatter returning to the sensor after the signal hits Earth's surface.For instance, vegetation and soil water content, soil roughness, and vegetation structure directly affect backscatter, while anthropogenic or climatic pressure can affect these four factors previously mentioned, indirectly affecting the backscatter (Vreugdenhil et al 2018, El Hajj et al 2019, Khabbazan et al 2019, Bueso et al 2023).However, among all cited mechanisms, in general, the dielectric properties of the medium are crucial in determining how much energy will be backscattered (Vreugdenhil et al 2020).Consequently, since the most important factor affecting vegetation's dielectric properties is moisture content, we can assume that backscatter increases with increasing VWC (Vreugdenhil et al 2020), and thus, SAR backscatter can be used to estimate VWC as already successfully demonstrated by previous studies (Vreugdenhil et al 2018, El Hajj et al 2019).However, similarly to any other remote sensing product, SAR-derived VWC estimates require proper validation.Previous studies validated VWC estimates either using other remote sensing products as reference or using field data acquired in croplands (Vreugdenhil et al 2018, 2020, El Hajj et al 2019).Thus, estimating VWC over natural ecosystems (such as savannas and shrublands) remains challenging due to their complex vegetation structure and diverse species composition, and therefore, validating a SARbased VWC product with field data acquired in these ecosystems could further demonstrate the potential of these data for VWC studies.
To address these challenges, we here aimed at investigating the potential of Sentinel-1 C-band SAR data for estimating VWC in natural ecosystems in central Brazil.Specifically, we propose to answer two main questions: (1) How well can Sentinel-1 SAR data capture variations in VWC over different vegetation types (i.e.dry and waterlogged grasslands, and savannas) and rainfall conditions?(2) How do the studied vegetation types respond to dry periods in terms of water content (e.g. which ones retain water better or recover water content faster)?It is noteworthy that here we focused on the seasonal dry periods, and not on anomalous drought events.To tackle these questions, we calibrated and validated VWC estimates derived from Sentinel-1 using as reference field sampled data in 82 plots in Central Brazil.After validation, we derived monthly VWC maps between April 2015 and September 2023 for grassland and savanna ecosystems within our study area.Finally, these maps were used to assess how the studied vegetation types respond to water shortage during the dry season in terms of community-scale water content.

Field data
The study was performed in the Chapada dos Veadeiros (CV), state of Goiás, central Brazil (figure 1), where 82 permanent plots of 10 × 10 m were established and VWC was sampled in April 2022.The plots were distributed in three distinct vegetation types, which are the dominant ones in the CV: dry grasslands (25 plots), waterlogged grasslands (25 plots), and Cerrado stricto sensu (32 plots).The latter can be broadly classified as a savanna vegetation type, and thus, hereafter will be referred to as 'savanna' .Further details about the study area and field plot distribution can be found in the supplementary material (appendix S1).
Besides the spatial variability, for the validation of a remote sensing product, it is also recommendable to consider the temporal variability in the studied biophysical parameter (in our case, VWC).This ensures that the reference values acquired in the field consist of a realistic representation of the VWC values found in the studied ecosystem.Thus, four field campaigns were performed in distinct seasons between April 2022 and February 2023: in the transition from the rainy to the dry season (April), in the dry season (August), in the transition from the dry to the rainy season (November), and in the rainy season (February).In each of these field campaigns, VWC at the plot-level was estimated through destructive sampling of grasses, leaves, and branches samples, and extrapolation using allometric equations; soil moisture was also measured in the field (more details about VWC estimation and soil moisture measurements can be found in appendix S1).

Remote sensing data
We used Ground Range Detected (GRD) Sentinel-1 C-band (5.405 GHz) SAR imagery, acquired at a 10 m spatial resolution in Interferometric Wide swath mode.Images were acquired from Google Earth Engine (GEE), which provides calibrated and ortho-corrected products.The downloaded GEE product was already pre-processed using the Sentinel-1 Toolbox through the following steps: orbit correction, GRD border noise and thermal noise removal, radiometric calibration, and terrain correction using the SRTM 30 digital elevation model (Filipponi 2019).However, since no speckle filtering and terrain normalization were performed in the GEE product, we further processed the data in the Toolbox using the framework proposed by Mullissa et al (2021): additional border noise correction (to deal with the border noise that remained from the GEE pre-processing), speckle filtering using the improved Lee sigma filter (Lee et al 2008), and radiometric terrain normalization using the SRTM 30 and following Hoekman and Reiche (2015).The code for these pre-processing steps was made available by Mullissa et al (2021).Both co-polarization (VV) and cross-polarization (VH) backscatter data were acquired, but because VH backscatter presented a better correlation with the reference field data, we focused here on presenting the results for VH backscatter.VH backscatter values were converted to decibels (dB) via log-scaling.Finally, to reduce noise in the VH imagery, we used both temporal and spatial aggregation: for the former, the mean value calculated from all available images in a month was used, aiming at representing the overall condition for that specific month; for the latter, images were aggregated to a 30 m spatial resolution using bilinear interpolation.
Rainfall data were used to illustrate how much water was available, monthly, when comparing water content changes through the seasons, and potential evapotranspiration (PET) data were used to illustrate the atmospheric demand for water in the interannual analysis (section 2.4).Monthly rainfall data were obtained from the Climate Hazards Infrared Precipitation with Stations (CHIRPS) dataset (Funk et al 2015), at 5 km spatial resolution.The average monthly rainfall value for the entire study area was used.PET data were obtained both 8-daily and annually, at a 500 m spatial resolution, from the MOD16A2 and MOD16A3GF products, respectively (Running et al 2017(Running et al , 2021)).Since the 8-daily PET data is provided as kg m −2 (8 d) −1 , values were divided by eight to obtain kg m −2 d −1 .The monthly average value, for the entire study area, was then used to represent the PET for each analyzed month.An overview of remote sensing data used can be found in table S1 in supplementary material.

VWC product derivation and error assessment
To calibrate VWC estimates from Sentinel-1 data, we employed a regression approach, correlating fieldbased VWC estimates with VH backscatter data (see figure S2 in supplementary material for an overview of the steps taken).Given that the relationship between VH backscatter and VWC is non-linear and saturates at higher backscatter values, we adopted an exponential model by log-transforming the response variable before fitting the regression (Vreugdenhil et al 2018).To explore the possibility of a more accurate representation of this non-linear relationship, we also experimented with a piecewise model ('segmented ' R package;Muggeo 2003).This model fits a regression with a segmented relationship, estimating the threshold in the independent variable that fits better with the data.Heteroskedasticity was assessed in both linear and piecewise models using White's test (White 1980), which indicated the absence of heteroskedasticity (P > 0.05).To account for seasonal effects on modeled VWC, we employed a Bayesian Generalized Model (BGM) including '(VH | season)' as a random-effect variable and 'sigma ∼ VH' to model the standard deviation as a function of the VH backscatter and consider heteroscedasticity ('brms' R package; Bürkner 2018).
To fit the models, all vegetation types were considered together.Field data from each field campaign was correlated with the VH backscatter mean value for that specific month (e.g.field data from April 2022 was modeled against the VH backscatter from April 2022, and so on).All models were trained on a structured-random subset comprising 80% of the observations for each vegetation type, while the remaining 20% constituted the validation set.Validation involved assessing VWC estimations and providing error estimates (we here used the root mean squared error, or RMSE).The use of an independent subset for validation allows for a robust assessment of the performance and intrinsic errors present in our approach (Hastie et al 2001).To identify the most suitable model for VWC estimation, we compared models' R 2 , Akaike Information Criterion (AIC), and RMSE (Burnham and Anderson 2004).For the BGM, Bayes-adjusted R 2 and the Widely Applicable Information Criterion were employed (Vehtari et al 2017, Gelman et al 2019).Finally, once the best model was determined, we re-fitted it using the complete dataset, as opposed to only the training subset.While this approach may impact the accuracy of error estimates, it maximizes the predictive capacity of the model by leveraging all available data (Roberts et al 2017).The model with the best performance (i.e. the piecewise model; see Results) was then used to estimate VWC from Sentinel-1 VH backscatter data (equation ( 1)): where VH is the VH backscatter signal.
We know that microwave radar data is sensitive to both vegetation and soil water content (Vreugdenhil et al 2018, Wang et al 2023).However, C-band's relatively short wavelengths limit its ability to penetrate densely vegetated ecosystems.In our study area, even vegetation types lacking woody species have a continuous herbaceous layer that attenuates C-band wave penetration into the soil.To assess the contribution of soil moisture on satellite-measured VH backscatter, we fitted a multiple regression using both VWC and soil moisture as explanatory variables and VH backscatter as the dependent variable.We included 239 data points in this analysis, excluding plots with zero soil moisture values or those affected by fire.Both explanatory variables were log-transformed, to improve normality, and scaled, to allow for coefficients comparison.This approach helps discern which variable predominantly influences VH backscatter measured by the satellite.

Vegetation response to the dry season
After obtaining a satisfactory model able to estimate VWC from VH backscatter data, we can use this model to estimate VWC for the entire study area and in distinct time steps within the Sentinel-1 time series (i.e. from April 2015 to September 2023).Since the models were adjusted using the natural logarithm of VWC (equation ( 1)), to estimate VWC values (in kg m −2 ) the exponential function was used to convert the values back from the log-scale.To better understand how each of the three vegetation types studied here responds to the water shortage during the dry season, in terms of VWC, we first focused on a short period, between December 2021 and February 2023, comprising one dry (June to August) and two rainy (December to February) seasons, and the respective transition seasons.Since the analysis was performed in a shorter period within the full time series and comprises one intra-annual growing cycle, hereafter we refer to it as the intra-annual analysis.However, since VWC is strongly dependent on vegetation biomass, normalization is necessary before comparing changes in water content among systems with distinct biomass levels.Thus, we used the Relative Water Content (RWC), which normalizes VWC using its dry and hydrated states, and is a classic indicator of plant water stress, integrating several plant physiological aspects related to drought response (Barrs andWeatherley 1962, Martinez-Vilalta et al 2019).RWC was calculated, pixel-wise, using the formula proposed by Rao et al (2019), which is an adaptation for the community scale of the conventional formula used for calculating RWC at the leaf level.In the conventional formula, the fresh weight of a leaf sample is subtracted by its dry weight, which is then divided by the difference between fully turgid and dry weight (Weatherley 1949, Barrs and Weatherley 1962).To adapt the formula to the community scale, using our VWC product, we followed what has been proposed by Rao et al (2019): the lowest and highest water content values for a given pixel, extracted from a long enough time series (we here used the full available time series), can be used to represent the typical dry and hydrated states of the plant community within that pixel.In our case, the full time series is composed of monthly VWC from April 2015 to September 2023.This allowed us to transfer the concept of RWC from a single leaf or plant to all plants within a community (or pixel), and then calculate RWC, for each pixel, based on equation ( 2 Next, based on the RWC time series, we were able to determine which vegetation type loses and regains water more rapidly in the transition seasons.This was done through a change analysis between consecutive time steps.That is, we used the differences between consecutive observations (i.e.deltas) to determine losses and gains of water content in the transition from the rainy to the dry season (March-May), during the dry season (June-August), and in the transition from the dry to the rainy season (September-November). Average delta values for each period were calculated, and this was done for each vegetation type separately, aiming at determining if certain vegetation types can buffer vegetation against water loss and/or rapidly recover water content.
To determine how communities from distinct vegetation types can retain water, even when atmospheric demand for water is high, we calculated a proxy for water holding capacity by dividing the total annual RWC by the annual PET.The former was obtained by calculating the area under the RWC curve (i.e. the total RWC during a given period) for each year in the time series; similar to what is done to estimate total growing season productivity in NDVI time series (e.g.Rasmussen 1992, Horion et al 2016, Bernardino et al 2020b).The latter was provided by the MOD16A3GF product (Running et al 2021).The resulting value represents the water holding capacity in m 2 y kg −1 , which cannot be interpreted meaningfully in biophysical terms, but can be used to compare the water holding capacity of distinct communities (and similarly, distinct vegetation types).Thus, we used the results to compare the capability of vegetation types to retain water by performing a Kruskal-Wallis followed by a Dunn's post hoc test for multiple comparisons.Since this analysis was performed for all the years where both RWC and PET data were available (i.e. from 2016 to 2021), we hereafter call it the inter-annual analysis.
Both intra and inter-annual analyses were performed for 100 randomly sampled pixels for each vegetation type, to reduce the impact of spatial autocorrelation on the results.The pixels were randomly selected in previously defined areas for each vegetation type, according to the land cover map (Lewis et al 2022), visual inspection of true-color satellite images, and field knowledge, to make sure that the pixels would be extracted from locations that truly represent the studied vegetation types.Additionally, monthly rainfall and PET data (see section 2.3) were used to illustrate how much water was available and what was the atmospheric demand in each month comprising the analyzed period.Finally, although the goal here was to evaluate the response of each vegetation type to the annual seasonal dry period, instead of extreme or severe drought events, we also briefly investigated how the RWC changed in periods with distinct drought conditions, according to the Standardized Precipitation Evapotranspiration Index (SPEI; Vicente-Serrano et al 2010, Beguería et al 2014).The mean time series of 20 pixels randomly sampled for each vegetation type were visualized together with the SPEI monthly time series for our study area (i.e. the average value, from each month, for all pixels within the study area).SPEI values below −2 were considered to be extremely dry (Vicente-Serrano et al 2010, Beguería et al 2014).Before randomly sampling the pixels, areas affected by a catastrophic fire that occurred in 2017 were masked, aiming at focusing on the changes caused by drought and disregarding the ones caused by this fire event.Moreover, RWC time series were de-seasonalized using the Seasonal Decomposition of Time Series by Loess (STL) in R (Cleveland et al 1990), allowing us to compare inter-annual changes in RWC regardless of seasonality.The RWC time series were clipped up to the end of 2022, to match the SPEI time series, which ends in 2022.Results are presented in figure S3 in the supplementary material.

Modeling VWC
The linear model between VWC estimated in the field and VH backscatter yielded an R 2 of 0.497 (P < 0.01), indicating that VH backscatter alone is already a good proxy for VWC.The BGM presented a similar R 2 , while the piecewise model performed slightly better both in terms of R 2 (0.520) and RMSE (0.465 kg m −2 ; table 1).The low RMSE for the piecewise model (i.e.0.465 kg m −2 , given that 95% of the VWC measurements made in the field ranged from 0.033 to 3.77 kg m −2 ) indicated that the errors in its predictions were lower, and thus, that it is better able to estimate VWC from unseen testing data when compared to the two other models (table 1).To illustrate the relationship between the plot-level VWC and the VH backscatter, a scatter plot with the fitted piecewise trend is presented in figure 2.
To compare the contribution of VWC and soil moisture on the VH backscatter measured by the satellite, we used a multiple regression including both as explanatory variables.This was done using a subset of the sampled data, as better explained in the Methods section.The multiple regression presented a higher R 2 (R 2 = 0.568; P < 0.01), as expected, than the simple linear regression of VWC over VH backscatter (R 2 = 0.45; P < 0.01).The regression coefficients for VWC (1.5) and soil moisture (0.88) Table 1.Performance of the three different models tested.Both the simple linear model (LM) and the piecewise model (PW) were fitted using only the VH backscatter as explanatory variable, while for the Bayesian Generalized Model (BGM) we used VH backscatter as the fixed effect and added the potential random effect of season.Performance metrics used to compare models were the R 2 (or the Bayes-adjusted R 2 for the BGM), AIC (or the Widely Applicable Information Criterion for the BGM), and the RMSE.indicate that the spatio-temporal variability in the VH backscatter signal is mainly due to changes in VWC rather than in soil moisture (i.e. the coefficient for VWC is almost double the coefficient for soil moisture).However, soil moisture still exerts a certain influence over the backscatter signal.This indicates that using soil moisture ancillary data could improve the derivation of VWC estimates, by disentangling the relative contribution of soil and vegetation over the measured backscatter signal.However, current soil moisture datasets have a very coarse spatial resolution (e.g. 25 km) (Sazib et al 2018, Dorigo et al 2019), hampering their use at the Sentinel-1 scale.Thus, we hereafter used the VH backscatter alone, retrieved from Sentinel-1 imagery, to estimate VWC over the entire study area and at different points in time (equation ( 1)).

Loss and recovery of water content in distinct vegetation types
Examples of the resulting VWC maps are presented in figure 3, for January and August 2022, during the rainy and dry seasons, respectively.In the rainy season, a west-east gradient is observed within the National Park borders, where higher VWC values are observed toward the east (figure 3(b)).This is due to the dominance of savannas in this region, while a mosaic of grasslands and savannas is present to the west.The same approach used to estimate VWC from Sentinel-1 SAR data for the maps in figure 3 was applied for the entire time series of images, from April 2015 to September 2023, resulting in monthly maps of VWC for this period (dataset: Bernardino et al 2024).This allowed us to investigate how water content decreases from the rainy to the dry season, and in contrast, how it recovers when rainfall returns in the subsequent rainy season.Additionally, we were able to estimate communities' water holding capacity, using the VWC dataset with ancillary PET data.However, as previously explained (section 2.4), instead of using absolute VWC values we used the RWC for these analyses, allowing for the comparison of communities with distinct levels of absolute VWC due to differences in biomass.In the intra-annual analysis, all vegetation types lost water during the dry season, as expected, but differences between vegetation types were observed (figure 4).The savannas were the communities that maintained the higher RWC levels during the dry season, while dry grasslands reached the lowest values (figure 4).Subtle differences were observed between the two grassland types: while waterlogged grasslands slowly lose water throughout the dry season, up until October, dry grasslands quickly reach the curve base around June, beyond which RWC values do not change much (figure 4).This was also observed in the change analysis (table 2), i.e. waterlogged grasslands keep losing water content  Savannas and waterlogged grasslands showed to be able to quickly recover water content when rainfall starts again, while dry grasslands took longer (table 2).The inter-annual analysis offered insights into vegetation types' water retention capabilities in relation to atmospheric demand (figure 5).Water holding capacity values were derived by dividing the total annual RWC by the annual PET (representing atmospheric water demand).Higher values indicate a greater capacity to retain water even when atmospheric demand is high.Across all years studied, savannas consistently exhibited significantly higher water holding capacity than both grassland types (figure 5).Additionally, in most analyzed years (five out of six), waterlogged grasslands displayed higher water holding capacity levels than dry grasslands (figure 5).In conjunction with water holding capacity, we plotted standardized rainfall data, calculated as the annual precipitation anomaly divided by the standard deviation (figure 5).We used the period from 1981 (start of the CHIRPS time series) to 2022 as the reference period for anomaly calculations.Overall, inter-annual variations in rainfall do not seem to impact the water holding capacity of the analyzed vegetation types.That is, rainfall deficit or surplus does not substantially alter water holding capacity.However, during years with the highest and lowest rainfall levels, distinctions among vegetation types were pronounced.In high-rainfall years, the most Values were calculated for each of the sampled plots (100 for each vegetation type).Comparisons are made within a year and letters represent significant (P < 0.05) differences between groups according to a Kruskal-Wallis followed by a Dunn's post hoc test with FDR correction.The standardized (stand.)rainfall is also presented, for the sake of comparison.notable contrast was observed between waterlogged and dry grasslands, while in low-rainfall years, it was between savannas and the two types of grasslands.Finally, we compared RWC time series for each vegetation type with the SPEI time series for the study area (figure S3).After several years with below-average rainfall levels, the area experienced two consecutive years with extreme drought conditions (2016 and 2017, where SPEI was below −2).This is reflected in the RWC time series by an abrupt drop in RWC in 2016, which remains low throughout 2017.After that, when conditions become less severe (2018 and 2019), a recovery in RWC is observed.

VWC estimation and uncertainties
In this study, we highlight the potential of Sentinel-1 SAR data for estimating VWC in natural ecosystems and monitoring changes in vegetation dynamics.The VH backscatter alone explained a large part of the variability in the field-measured plot-level VWC (R 2 = 0.52; P < 0.05).Another study, performed on agricultural fields of distinct crop types, also found a strong relationship between corn VWC and VH backscatter using an exponential model (R 2 = 0.62; P < 0.05), but for other crop types, such as oilseedrape and winter cereal, the relationship was not significant (Vreugdenhil et al 2020).In addition, Vreugdenhil et al (2020) used reference data from croplands, while to our knowledge, our study is the first one to provide VWC estimates from Sentinel-1 data validated over three distinct natural vegetation types.Here, we also found that SAR-derived VWC estimates can be improved by better disentangling the relative contribution of soil water and VWC over the SAR signal (Vreugdenhil et al 2020, Wang et al 2023), given that soil moisture directly influences VH backscatter (although more subtly, when compared to VWC).
Using the complete Sentinel-1 time series, from April 2015 to September 2023, and the fitted piecewise model, we generated monthly VWC maps with a spatial resolution of 30 meters for the CV study area in Central Brazil.The development of such a product with a relatively high spatial resolution is valuable for Cerrado regions as well as other regions globally that present a mosaic of distinct vegetation types within a small land area.For instance, at the MODIS resolution (250-500 m), in the CV study area, it is not rare to find two to three distinct vegetation types with large variations in VWC.In the generated VWC maps, regions with higher water content levels may exhibit larger errors in VWC estimates due to a saturation effect observed in the backscatter signal (figure 2(a)), and thus, changes in VWC might not be accurately reflected by variations in VH backscatter in these areas.However, our validation process enabled us to assess the inherent errors in the VWC estimates, which were relatively modest (i.e.RMSE of 0.465 kg m −2 ; table 1).This level of error is considered low, given the wide range of VWC values recorded in the field (ranging from 0.033 to 3.77 kg m −2 in 95% of the measurements).In comparison, a study using optical remote sensing data for VWC estimation in agricultural fields achieved an RMSE of 0.576 kg m −2 for corn and 0.171 kg m −2 for soybean (Jackson et al 2004).However, it is noteworthy that such estimates in natural ecosystems can be more challenging due to the complexity and variability in vegetation structure within these systems.

Vegetation response to seasonal dry periods and evaporative demand
Seasonal variations in vegetation RWC are evident, particularly during the dry season (figures 3 and 4).In grasslands, RWC drops to 0.015 and 0.032 (on average) for dry and waterlogged grasslands, respectively (figure 4).This large drop in RWC during the dry season, primarily in dry grasslands, arises from the grasses' water-use strategy and phenological behavior.They rely on the sporadic water available in shallow soil layers and wither aboveground when water is scarce (Burgess 1995, Scanlon et al 2002).In contrast, savannas maintain RWC above 0.101, likely due to woody vegetation's water-retention abilities or access to deeper water sources (Holbrook 1995, Oliveira et al 2005).Dry grasslands quickly reach their RWC minimum at the dry season's onset, while waterlogged grasslands gradually lose water content throughout (figure 4 and table 2).Waterlogged grasslands, closer to the water table, can delay water loss during the middry season, but they also struggle to maintain RWC levels during the initial dry months (February-May).Here, it is noteworthy that we used a single year to assess the intra-annual response of vegetation, and thus, results might be affected by atypical seasonal conditions.However, 2022 was selected as the year for this analysis since it presented close to normal rainfall conditions, according to the SPEI data (figure S3; SPEI of −0.35 on average), among the years overlapping with the Sentinel-1 time series.Therefore, we believe that no major climatic anomaly has affected these results.
Inter-annual analysis reveals that savannas consistently hold more water than both grassland types in all years (figure 5).This difference is likely due to niche complementarity where woody and grassy vegetation coexist: savannas' woody component retains water better and deep-rooted individuals reach deeper groundwater, while grasslands rely on the water present in shallower soil layers (Holbrook 1995, Oliveira et al 2005).This finding aligns with the intra-annual analysis, but here evidencing that what was found is consistent throughout different years and regardless of inter-annual changes in rainfall.Waterlogged grasslands outperform dry grasslands in water holding capacity in most years because they have prolonged access to the water table (year-round in some communities).In 2016, the year with the lowest rainfall between 2016 and 2021, savannas exhibited the largest differences in water holding capacity compared to grasslands (figure 5).This suggests a decoupling between savannas' water retention capabilities and rainfall availability, warranting further investigation.

Future perspectives and limitations
We here focused on natural systems in the Cerrado domain, Central Brazil.The Cerrado harbors the headwaters of most watersheds in Brazil and is fundamental for the maintenance of water quality in the Brazilian river system (Lima andSilva 2007, Hunke et al 2015).It is also an important provider of ecosystem services for local communities (Resende et al 2019(Resende et al , 2021)).Although this highly enriches the importance of our findings, focusing on the local scale means that further studies are needed before our approach can be applied to different vegetation types and/or different regions.Thus, future studies could aim at calibrating and validating VWC estimates from Sentinel-1 data in distinct ecosystems, contributing to the evaluation of the potential of SAR-derived VWC estimates.
Although expanding our results for other vegetation types and regions around the globe is important, the VWC maps derived here, limited to the Chapada dos Veadeiros study area, are useful for several studies being developed in the region.For instance, they can be used for further analysis of ecosystem dynamics, including ecosystems' response to drought events and VWC-based thresholds that lead to increased mortality or loss of productivity, which ultimately can support land management decisions.

Figure 1 .
Figure 1.Study area and distribution of the field plots.The study was performed in Chapada dos Veadeiros, central Brazil (yellow dot in the country-level map).The Chapada dos Veadeiros National Park borders are highlighted in yellow.Inset maps show a closer view of the five areas where the field plots were established.Photos illustrating the vegetation types sampled are presented in figure S1 in the supplementary material.Background imagery provided by Google Earth (2018).

Figure 2 .
Figure 2. Relationship between VWC, soil moisture, and the VH backscatter.In (a), the relationship between total plot VWC and VH backscatter is presented (dashed lines represent the fitted piecewise trend), using different colors to represent the different seasons when field data was sampled ('transition 1' represents the transition from rainy to dry, and 'transition 2' the transition from dry to rainy).The full dataset was used, instead of using only the training data.The relationship between the VH backscatter and both VWC (b) and soil moisture (c) are also presented.Although (b) is almost a repetition of (a), the dependent variable in the former is VH backscatter, since the purpose of (b) and (c) is to visualize the contribution of these variables to the backscattered signal.Moreover, the dataset used is slightly different, since data points where soil moisture was zero were excluded (as specified in the Methods).

Figure 3 .
Figure 3. Land cover and VWC maps for the study area.(a) Land cover map of the study area.Reproduced from Lewis et al (2022).CC BY 4.0.Please note that savannas are regionally called 'Cerrado stricto sensu' or 'Cerrado rupestre' , depending on the substrate where they occur (i.e. the latter occur on rocky substrates).The VWC maps presented are for January 2022 (b) and August 2022 (c), during the rainy and dry season, respectively.The same color scheme was used to allow for the comparison between maps.

Figure 4 .
Figure 4. Temporal changes in RWC for one year and five months, from December 2021 to February 2023.RWC dynamics are presented for each vegetation type separately, with the lines representing the average value for the 100 sampled pixels.Changes in RWC are compared to monthly rainfall and PET.

Figure 5 .
Figure 5. Water holding capacity for the three studied vegetation types, between 2016 and 2021.The water holding capacity indicates how much water the community can hold based on the atmospheric demand (see Methods).Values were calculated for each of the sampled plots (100 for each vegetation type).Comparisons are made within a year and letters represent significant (P < 0.05) differences between groups according to a Kruskal-Wallis followed by a Dunn's post hoc test with FDR correction.The standardized (stand.)rainfall is also presented, for the sake of comparison.
RWC t = VWC t − VWC min VWC max − VWC min (2) where t represents each time step in the time series; VWC min represents the lowest and VWC max the highest VWC value extracted from the full time series, from April 2015 to September 2023. ):

Table 2 .
RWC change analysis between seasons in 2022.The analysis comprised a transition season from rainy to dry, a dry season, and the consecutive transition from dry to rainy.Positive values represent RWC gain, while negative values represent RWC loss.