The model used in this study is LHF (Fan et al., 2013; Miguez-Macho and Fan, 2012a, b; Pokhrel et al., 2013, 2014), a continental-scale land hydrology model that resolves various land surface hydrologic and groundwater processes on a full physical basis. It is derived from the model Land Ecosystem–Atmosphere Feedback (LEAF) (Walko et al., 2000), the land surface component of the Regional Atmosphere Modeling System (RAMS) (Pielke et al., 1992). The original LEAF was extensively improved and enhanced to develop LEAF-Hydro for North America (Fan et al., 2007; Miguez-Macho et al., 2007) by adding a prognostic groundwater storage and allowing (1) the water table to rise and fall or the vadose zone to shrink or grow; (2) the water table, recharged by soil drainage, to relax through streamflow into rivers, and lateral groundwater flow, leading to convergence to low valleys; (3) two-way exchange between groundwater and rivers, representing both losing and gaining streams; (4) river routing to the ocean as kinematic waves; and (5) setting sea level as the groundwater head boundary condition. Miguez-Macho and Fan (2012a) further enhanced the LEAF-Hydro framework by incorporating the river–floodplain routing scheme which solves the full momentum equation of open-channel flow, giving more realistic streamflow estimates by considering the prominent backwater effect observed in the Amazon (Bates et al., 2010; Yamazaki et al., 2011). The LHF model has been extensively validated in the North and South American continents on 5 and 2 km grids, respectively (Fan et al., 2013; Miguez-Macho et al., 2008; Miguez-Macho and Fan, 2012a, b; Pokhrel et al., 2013; Shin et al., 2018) and used to examine the impacts of climate change on the groundwater system in the Amazon (Pokhrel et al., 2014). A complete description of LHF can be found in Miguez-Macho and Fan (2012a).

Atmospheric forcing data are taken from WATCH Forcing Data methodology applied to ERA-Interim reanalysis data (WFDEI) (Weedon et al., 2014), available for the 1979–2016 period at 0.5∘ spatial resolution and 3 h time steps. The WFDEI dataset is widely used in both global- and regional-scale studies (Beck et al., 2016; Felfelani et al., 2017; Hanasaki et al., 2018; Müller Schmied et al., 2014) and has been suggested to represent the observations in the Amazon region well (Monteiro et al., 2016). The original WFDEI data at 0.5∘ resolution are spatially interpolated using a bilinear interpolation method to model grid resolution (∼2 km), following our previous studies (Miguez-Macho and Fan, 2012a, b; Pokhrel et al., 2013, 2014; Shin et al., 2018). The more recent European Centre for Medium-Range Weather Forecasts Reanalysis 5th (ERA5) dataset, which provides atmospheric forcing data from 1979 to present day at a spatial resolution of 0.25∘, shows promise by outperforming its predecessors (Towner et al., 2019). However, as no studies existed in the past literature which comprehensively validated the ERA5 dataset over the Amazon region until recently, WFDEI forcing remains a better alternative as a model input.

The land cover data used in this study are obtained from the European Space Agency Climate Change Initiative's Land Cover project (ESA-CCI; http://maps.elie.ucl.ac.be/CCI/, last access: 24 June 2019). The data comprise an annual time series of high-resolution land cover maps for the 1992–2015 period at a 300 m spatial resolution, generated by combining the baseline map from the Medium-spectral Resolution Imaging Spectrometer (MERIS) instrument and the land use–land cover (LULC) changes detected from AVHRR (1992–1999), SPOT-Vegetation (1999–2012), and PROBA-V (2013–2015) instruments. The classification follows the LULC classes defined by the UN Land Cover Classification System (LCCS). Spatiotemporal coverage and resolution of these LULC maps are consistent with the specific LHF model requirements; hence we use annual land cover input, spatially aggregated to 2 km LHF model grids, following the general practice in hydrologic impact studies (Arantes et al., 2016; Panday et al., 2015).

Because the ESA-CCI data did not cover the simulation period prior to year 1992, we derive the time series products for the 1980–1991 period by using the trend in leaf area index (LAI) and the ESA-CCI land cover map for year 1992 as a baseline. A pixel-by-pixel analysis is conducted, and the pixels with mean annual LAI higher than 5 are transitioned into the forest canopy, whereas for other pixels the LULC type is retained from the previous year's LULC map. The threshold of LAI equal to 5 for facilitating the land cover transition into forest is determined based on the LAI classifications provided in past literature (Asner et al., 2003; Myneni et al., 2007; Xu et al., 2018). Reverse prediction of LULC changes was constrained to the forest canopy only, as it is difficult to predict the LULC type based on LAI values less than 5. Also, forest cover is known to be the most prominent land cover in the Amazon; hence it is reasonable to assume that most of the LULC changes occurring in the basin are transitioned from forest cover.

Monthly LAI data are derived by temporally aggregating the 8 d composites from the Global Land Surface Satellite (GLASS) LAI product (Liang and Xiao, 2012; Xiao et al., 2014) to monthly values for the entire model domain. GLASS LAI values for the period of 1982–1999 are derived from AVHRR reflectance, whereas MODIS reflectance values are used for the period 2000–2012. Because of the data constraint, LAI data for the years before 1982 and after 2012 are assumed to be the same as that of the years 1982 and 2012, respectively.

To examine the occurrence and severity of hydrological droughts over the past decades, we employ the drought severity index derived from time-varying TWS change from GRACE, known as the GRACE drought severity index (GRACE-DSI) (Zhao et al., 2017b). We apply the GRACE-DSI framework to the 36-year simulated TWS (hereafter referred to as TWS-DSI) to examine the interannual and interdecadal drought evolution over the entire basin. This index is solely based on the TWS anomalies and has been shown to capture the past droughts with favorable agreement with other drought indices derived from precipitation (e.g., PDSI and SPEI) (Zhao et al., 2017a, b). TWS-DSI is calculated for each grid cell in the model domain as follows: 1TWS_DSIi,j=TWSi,j-TWS‾jσj, where TWSi,j is the TWS anomaly from LHF for year i and month j; and TWS‾j and σj are the temporal mean and standard deviation of TWS anomalies for month j, respectively.

The characteristics of hydrological droughts are identified from the simulated streamflow using the widely used threshold level approach. Different thresholds have been proposed in previous studies: mean flow, minimum and maximum flows (Marengo and Espinoza, 2016; Wongchuig Correa et al., 2017), 80th percentile (Q80) flow (Van Loon et al., 2012; Van Loon and Laaha, 2015; Wanders and Van Lanen, 2015), and 90th percentile (Q90) flow (Wanders et al., 2015; Wanders and Wada, 2015). In this study, we use Q90, which is derived from the flow duration curve, where Q90 is the streamflow that is equaled or exceeded for 90 % of the time. Q90 is used to isolate severe drought events over the simulation period. Monthly threshold values are derived using the 36-year simulated streamflow and are smoothed by a 30 d moving average. Drought condition is identified by determining whether the variable is below the threshold, expressed mathematically as 2Ds(t,x)=1forQ(t,x)<Q90(t,x)0forQ(t,x)≥Q90(t,x), where Ds(t,x) indicates whether the grid (x) is in a drought state at time (t), Q(t,x) is the streamflow, and Q90(tx) is the threshold for grid (x) at time (t). Consecutive drought states are added to get the drought duration. Events with duration less than 3 d are not considered as droughts. The number of drought days per year is calculated by aggregating the duration of all the drought events in a year.

We define the dry-season total water deficit (TWD) as the cumulative difference between monthly potential evapotranspiration (PET) and precipitation (P) for the period during which P<PET. The corresponding drop in the simulated TWS, during the same period as of TWD, is defined as the TWS release (TWS-R). TWD and TWS-R can be conceptualized as the annual water demand and supply as described in Guan et al. (2015). PET estimated at the daily interval using the Penman–Monteith approach (Monteith, 1965) as in Pokhrel et al. (2014) is aggregated to the monthly scale to calculate TWD; for consistency, we use the WFDEI forcing data that are used for LHF simulations (Sect. 2.2). TWS anomalies required for the estimation of storage release are obtained from the LHF model.

LHF is set up for the entire Amazon basin (∼7.1 million square kilometers) including the Tocantins River basin. Simulations are conducted for the 1979–2015 period at a spatial resolution of 1 arcmin (∼2 km). The model time step is 4 min as in previous studies (Miguez-Macho and Fan, 2012a, b; Pokhrel et al., 2013, 2014); however, model output is saved at daily time steps. To stabilize water table depth, the model is spun up for ∼150 years, starting with the equilibrium water table (Fan et al., 2013) for 1979, and results for the 1980–2015 period (36 years) are analyzed. As this study aims to analyze the hydrological changes in the Amazon on a decadal scale, simulations for 1979 are considered as additional spin-up and hence not used. Dynamic monthly LAI and annual LULC maps are used to account for LULC changes (see Sect. 2.3). Moreover, as the model simulates land surface, hydrologic, and groundwater processes on a complete physical basis, no calibration was performed on the model output. The original novelty of the LHF model framework, combined with the incorporated dynamic human role through land cover change, creates a state-of-the-art framework for assessing long-term hydrological changes. The complete LHF framework along with the input data employed in this study are presented in Fig. S1 in the Supplement.

Figure 1 presents the Taylor diagram (Taylor, 2001) illustrating the statistics of the simulated streamflow against observations at 55 gauging locations (see Sect. 2.4.1 and Fig. S2) across the entire Amazon basin. The Taylor diagram provides a synthetic view of error in the simulations in terms of the ratio of standard deviation (SD) of the simulated streamflow to the observed as a radial distance and their correlation as an angle in the polar axis. Most of the stations show a high correlation (>0.8) and a SD ratio close to unity, indicating a good model performance overall for varying geographical locations and stream sizes over the Amazon. Low correlation (∼0.6) is seen for some gauging stations situated on streams with smaller annual mean flow and steep slope profile, for example, the smaller streams across the Andes in Japura and Negro sub-basins, along with the streams in the northeastern parts of the Amazon. In these streams with high topographic gradients, precipitated water quickly flows away, causing slightly erratic patterns of seasonal streamflow, which is apparent in both simulated and observed time series (Figs. S3 and S4). However, due to the difficulty in resolving hillslope processes for low-order streams using 2 km grids, the model is unable to fully capture the flow seasonality in the streams with high topographic gradient.

The spatial distribution of the simulated streamflow across the entire model domain and the time series comparison of simulated vs. observed streamflow at 12 selected stations are presented in the Supplement (Figs. S2 and S3). The simulated seasonal cycle compares well with the observed one for the entire basin (i.e., Obidos station) as well as for most sub-basins; however, discrepancies in the seasonal peaks can be seen in some basins (e.g., Xingu, Tocantins, and Tapajos). Man-made reservoirs generally attenuate streamflow peaks and seasonal variability, reducing the SD, which is reflected in the observed data but not yet accounted for in the model; this could have exaggerated the SD ratio in some cases. For example, the streamflow in the Tocantins River shows higher SD compared to observed streamflow, likely due to the operation of the Tucurui I and II dams. Conversely, the SD ratio is lower than unity at some stations, including those on the Madeira River (Fig. 1), due to the dry bias found in the input precipitation (see Fig. S5 and Sect. 3.2). For sub-basins with higher groundwater contribution to streamflow, such as Xingu, Tapajos, Tocantins, and Madeira, the dry-season flow is overestimated (Fig. S3), which results from possibly exaggerated groundwater buffer in the model for these regions (Miguez-Macho and Fan, 2012a). Given that LHF is a continental-scale model, simulates streamflow on a full physical basis, and is not calibrated with observed streamflow, we consider these results to be satisfactory to study the hydrologic changes and variability.

Figure 2 presents the comparison of simulated TWS anomalies and GRACE data for the entire Amazon basin and its eight sub-basins; for model results, the individual TWS components are also provided. The model performs very well in simulating the basin-averaged TWS anomalies for the entire Amazon basin and most sub-basins. However, some differences between the simulated and GRACE-based TWS anomaly are evident, especially in sub-basins with a relatively smaller area and elongated shape (e.g., Purus and Japura). Note that the accuracy of GRACE–model agreement is generally low in such small basins due to high bias and leakage correction errors (Chaudhari et al., 2018; Felfelani et al., 2017; Longuevergne et al., 2010), reflected by higher root mean square error (RMSE) values in Fig. 2. Simulated TWS evidently follows precipitation anomalies (shown in grey bars in Fig. 2), implying that any uncertainties in the precipitation forcing could have directly impacted TWS. For example, the simulated TWS peak in 2002 in the Solimoes River basin results from the anomalous high precipitation; however this could not be validated due to a data gap in GRACE. Overall, the model performance is better in the first half of the simulation period (i.e., 2002–2008) compared to the second half, especially in the western sub-basins including the Solimoes and Japura, which could be partially attributed to the decreasing trend in the precipitation forcing noted in Fig. S6.

Figure 2 also shows the seasonal cycle including the contribution of different storage components to TWS. In all the basins, the simulated seasonal cycle matches extremely well with GRACE, adding more confidence to the model results. TWS signal is sturdily modulated by the sub-surface water storage, demonstrating the importance of groundwater in the Amazon, especially in the southwestern sub-basins. The inverse relationship in the seasonal cycle of two sub-surface water stores, viz. soil moisture and groundwater, is readily discernable in Fig. 2, which is caused by the competing use of the sub-surface compartment by the two terms (Felfelani et al., 2017; Pokhrel et al., 2013). However, in some sub-basins, such as the Purus, Solimoes, and Negro, the low-lying areas with large floodplains cause floodwater storage to be equally prominent.

Here, we present a more detailed examination of the simulated TWS by comparing its spatial variability and trend with GRACE data. Because a shift in agreement between model and GRACE was detected in Figs. 2 and S7, we conduct a trend analysis for two different time windows: 2002–2008 and 2009–2015 (Fig. 3). It is evident from Fig. 3 that the model captures the general spatial pattern of TWS trend in GRACE and its north–south and east–west gradients especially for the first half of the analysis period; however, notable differences are evident in the second half (2009–2015), particularly over the Madeira River basin. This is a noteworthy observation given that the basin-averaged TWS variability matches extremely well with GRACE data (Fig. 2) and thus warrants further investigation. There could be a number of factors contributing to the disagreement, some of which could be model-specific (e.g., wet bias in simulated discharge; Fig. S3); however, this is a general pattern observed in many hydrological models as reported in a recent study (Scanlon et al., 2018).

Scanlon et al. (2018) indicated a low correlation between GRACE and models, which they attributed to the (i) lack of surface water and groundwater storage components in most of the models, (ii) uncertainty in climate forcing, and (iii) poor representation of human intervention in the models (Scanlon et al., 2018; Sun et al., 2019). Here, we shed more light on the disagreement issue by investigating the contributions from the explicitly simulated surface and sub-surface storage components and their latitudinal patterns, addressing the first concern noted above which is the most critical among the three in the Amazon because of the varying contribution of different stores across scales (Pokhrel et al., 2013). Figure 4 shows trends in TWS anomalies from GRACE products and the LHF simulation for the complete model–GRACE overlap period (i.e., 2002–2015) with climatology and with climatology removed; for LHF results, the surface and sub-surface component contributions to the TWS are shown. Also shown in the figure are the zonal means.

Simulated TWS from the LHF model displays a higher correlation with GRACE trends compared to most of the global models discussed in Scanlon et al. (2018). Due to the incorporation of a groundwater scheme and other surface water dynamics, the trend in basin-averaged TWS with climatology removed for the Amazon River basin is found to be -1.64 mm yr-1, much less negative than most of the simulated TWS trends reported in Scanlon et al. (2018). The difference in the sign of trend can partly be explained by the negative trend observed in the WFDEI precipitation (Fig. S6), concentrated over the Andes region which eventually drains into the main stem of the Amazon through the Solimoes River. Due to steep topography, the impact of decreased precipitation over the Andes range is carried over to its foothills in terms of runoff, hence corresponding well with the negative trends in simulated surface water storage over the Central Amazon (Fig. 4). Lower recharge rates in the region with decreasing precipitation trend (Fig. S6) are also very likely, which is supported by the negative trend visible in the sub-surface water storage in Fig. 4, over the northwest region of the Amazon. Hence, it can be concluded that, even though the model shows some bias in TWS compared to GRACE data, the model accurately represents the key hydrologic processes in the Amazon basin; yet, these results should be interpreted with some caution while acknowledging the uncertainty in the forcing dataset. We also emphasize that it is important to evaluate models using spatiotemporal trends, especially with GRACE, instead of just using the basin-averaged time series, a commonly used approach in most previous studies.

Figure 5 show the interdecadal shifts in mean simulated TWS (total and its components) for the simulation period. Several observations can be made from this figure. First, the change between the 2010s and 2000s suggests high negative anomalies in all the water stores, especially over Central Amazon. This is likely a result of increasing drought occurrence and severity in the region, e.g., the 2010 (Lewis et al., 2011; Marengo et al., 2011) and 2015 (Jiménez-Muñoz et al., 2016) Amazonian droughts. Second, although the 2000s encompassed one of the severe Amazonian droughts, viz. 2005 (Marengo et al., 2008; Zeng et al., 2008), its impact was not pronounced in terms of the decadal mean, which could be due to the offset caused by anomalous wet years including 2006 and 2009 (Chen et al., 2010; Filizola et al., 2014). Third, we find an increase in river water storage in the northwestern region and decrease in the southwest of the Amazon on a decadal scale (Fig. 5, column 1, row 2), which is in line with the findings reported in previous studies based on the observed streamflow in 18 sub-basins for the 1974–2004 period (Espinoza et al., 2009; Wongchuig Correa et al., 2017).

The most remarkable feature we observe in Fig. 5 is the exceptional interdecadal shifts between the 2000s and 2010s. The central and northwestern part of the Amazon region, encompassing the Negro and Solimoes, along with some parts of the Madeira in the southwest, experienced a major decadal dry spell compared to the previous decades. Although a major part of this decadal dry condition could be attributed to the decreasing trend in input precipitation discussed in Sect. 3.3 (Fig. S6), the regional hydrologic changes in terms of TWS are also prominent. Another peculiar phenomenon observed at the decadal scale is the start of the negative anomaly in groundwater storage over the Central Amazon. A small but spatially well distributed below-decadal-average water table (dictated by groundwater storage) is evident in the Central Amazon region and the upper stretches of the Madeira basin during the 2010s (Fig. 5, column 3, row 4). Since the water table is shallow, and groundwater is the major contributor of streamflow in this region (Miguez-Macho and Fan, 2012a), some part of the negative anomaly in surface water stores can be attributed to the below-decadal-average groundwater table.

Significant long-term trends in simulated TWS and its components are evident in sizeable portions of the basin (Fig. 6). While a negative trend is found in the southern and southeastern regions (e.g., Madeira, Tapajos, Xingu and Tocantins), the trend is positive in the northern and western regions (Solimoes and Negro) (see Fig. S9 for basin-averaged trends). Being the major contributor, sub-surface water storage mimics the trend patterns in TWS (see Sect. 3.2). On the contrary, surface water storage trends are mainly dominated by floodwater and are concentrated along the main stem of the Amazon and the upper reaches of the Negro. The positive trends in floodwater can be explained by the corresponding trends in input precipitation (Fig. S5). Excess precipitation in sub-basins, such as the Solimoes and Negro, which are characterized by a high topographic gradient, is directly translated in the surface water storage, in this case floodwater. Although a corresponding increment in river water storage is also expected, its smaller storage makes the trend magnitudes negligible. Nominal negative trends, but significant, in floodwater storage are found in the upper reaches of Madeira as well, corresponding to the negative trends in input precipitation over that region.

To provide an in-depth understanding of the interdecadal changes occurring in the Amazon region and to determine whether the changes observed in Fig. 5 are significant, we applied a t test methodology to the long-term TWS anomalies at basin and sub-basin levels. The spatial changes observed in Fig. 5 are summarized with their interdecadal significance in Table S1 (Supplement), along with the decadal means and standard deviations. Significant change at 99 % level is found in the Negro River basin throughout the study period, followed by the Solimoes River basin exhibiting significant change in the last three decades. These changes can be attributed to the corresponding changes in precipitation (Fig. S5), which follow a similar change in respective basins. However, the significant hydrologic changes in the Tocantins and Madeira can be primarily attributed to LULC changes, as the corresponding changes in precipitation were relatively negligible. For example, the Tocantins River basin underwent major LULC changes in response to heavy deforestation caused by dam construction and cattle farming (Costa et al., 2003) until policies were imposed in 2004 by the Brazilian government (captured in the ESA dataset; Fig. S8). Similarly, the Madeira River basin also endured major LULC changes in the late 1990s, which were dominated by agricultural expansion (Dórea and Barbosa, 2007).

As a first attempt to comprehensively characterize the Amazonian droughts, we present a summary of all the drought characteristics discussed in the previous sections on a spider plot (Fig. 10). Each spider plot is a representation of a drought year with respect to the (i) causes of drought and their type in terms of common indices; (ii) response of different water stores, such as TWS, to the drought event; (iii) role of groundwater storage in alleviating the dry conditions on the surface; and (iv) the spatial impact of the drought in different sub-basins of the Amazon. Although no significant trend in the combined drought characteristic is apparent, Fig. 10 provides important insights into the variability of Amazonian droughts. It is evident from the figure that the drought variability over the years was significant in terms of both magnitude and spatial impact. The most notable feature in Fig. 10 is the distinct relationship between the SPI and drought duration. For example, during the 1995 drought, most of the river basins (e.g., Tocantins, Tapajos, Xingu, and Negro) experienced significant meteorological and TWS droughts; however, the severity of hydrological droughts was relatively negligible in those basins. Groundwater–surface water exchange is the key mechanism behind this unique behavior, causing groundwater to fulfill the drought deficit in streamflow over the basin. Due to shallow water tables at the downstream end of these basins, a significant quantity of groundwater is fed to the rivers, which manifests as high peaks in total groundwater release, evident in Fig. 10. Similarly, a high number of drought days are found corresponding to less groundwater release, such as during the 1995 drought in Madeira. Conversely, TWS-DSI generally follows the same pattern as that of the SPI but with a lesser magnitude, which can be attributed to the delayed response from groundwater.

Further, the behavior of the Amazonian sub-basins can be characterized by the shape of the polygon formed by the comparison of different aspects of past droughts. The convex and concave characteristic in the plots mainly depends on the interrelation between meteorological and hydrological drought indices, which is further controlled by the sub-surface water storage. A convex polygon indicates a lower groundwater contribution to streamflow in the sub-basin, such as in Purus during 1995 and 2005, whereas a concave polygon suggests higher groundwater release to streamflow in that particular year.

Results suggest an increasing trend in TWD with significant decadal variability over the Amazon and its sub-basins, indicating an increase in dry-season length over the past 36 years (Fig. 11). Further, the increasing gap between TWD and TWS-R suggests an intensified terrestrial hydrologic system over the dry season during the study period. As the LULC impact is partly accounted for in the PET calculations (i.e., through changing surface albedo), the river basins with substantial LULC change, such as Madeira, Tapajos, Tocantins, and Xingu, portray higher TWD trend magnitudes (significance >95 %). The peaks in the TWD correspond well with drought years; for example, the peaks in the TWD for Madeira are analogous to the drought years (e.g., 1988, 1995, 2005, and 2010). Due to this definitive response to drought conditions, TWD is also used to characterize historical drought events in earlier sections. We note that the trends in the total deficit should be interpreted with caution as the uncertainty in the forcing could have affected TWD and TWS-R trend estimates.

We find that the river basins that contain high altitudinal areas (Purus, Solimoes, and Negro) have a fairly balanced relationship between TWD and TWS-R, but southern and southeastern sub-basins exhibit a higher water deficiency (Fig. S11), with approximately 2- to 3-fold differences between TWD and TWS-R during regular years. For drought years, however, the difference between TWD and TWS-R is even higher, creating highly anomalous dry conditions in the sub-basins. Consistent higher values of TWD in southern and southeastern sub-basins of the Amazon further highlight the intensification of the dry season, with increasing water deficiency corresponding to an almost constant water supply from TWS-R. This phenomenon is also highlighted in Espinoza et al. (2016), which showed an significant increase in dry day frequency in the central and southern parts of the Amazon. Results from this study combined with the reported increasing trend in the wet season (Gloor et al., 2013) imply an overall intensification of the Amazonian hydrological cycle.