Our research employed the GRACE Mascon solutions provided by CSR, encompassing monthly Terrestrial Water Storage Anomaly (TWSA) sequences from April 2002 to November 2025. The reflected TWSA were measured relative to the average baseline spanning from 2004 to 2009. Prior studies have indicated that Mascon solutions present substantial enhancements over spherical harmonics in terms of leakage errors and post-processing (Scanlon et al., 2016). They do not necessitate additional smoothing, empirical de-striping, filtering, or scaling factors, and are more user-friendly (Save et al., 2016; Scanlon et al., 2016). The newly released RL0603 Mascon solutions adopt a novel grid structure founded on equidistant geodetic grids with a spatial resolution of 1°, which is resampled to a resolution of 0.25° for publication. This effectively mitigates the leakage between land and ocean signals. In this investigation, we utilized SSA to fill the 35-month data gaps. In combination with the GLDAS, we constructed the GWSA sequence. Moreover, we applied machine learning and deep learning methods to downscale the spatial resolution to 0.008333° (approximately 1 km at the equator), with the aim of attaining globally seamless high-resolution GWSA mapping with spatiotemporal resolution.

To assess the influence of different GRACE products on the GWSA downscaling model, we further introduced the JPL GRACE/GRACE-FO RL06.3Mv04 CRI-filtered mascon product for comparison with CSR. The JPL lwe_thickness variable represents TWSA in centimetres of equivalent water thickness and has a sampling resolution of 0.5°. We bilinearly resampled the monthly JPL grids to the CSR reference grid and used the common period from April 2002 to December 2020; missing months during the GRACE mission period were retained during model training (Scanlon et al., 2016; Wiese et al., 2016).

NASA GLDAS is designed to integrate satellite and ground-based observational data. By leveraging land surface modeling and data assimilation techniques, it aims to attain optimal simulations of land surface conditions and flux fields (Rodell et al., 2004). It encompasses three land surface models: the Noah model, the Catchment (CLSM) model, and the Variable Infiltration Capacity (VIC) model.

The Noah land surface model includes four soil temperature and moisture layers, along with models for vegetation canopy, snow cover prediction, ET, and soil drainage runoff. Owing to its precise representation of surface energy and water budgets, it has been extensively employed in global hydrological and water resource change research (Ek et al., 2003; Niu et al., 2011; Rodell et al., 2004).

In this study, monthly soil water storage (SWS), snow water equivalent (SWE), and vegetation canopy water storage (CWS) from GLDAS Noah V2.1 were used for the GWSA inversion. The soil-water-storage term was calculated as the vertically integrated storage of the four Noah soil layers, using SoilMoi0_10cm_inst (0–10 cm), SoilMoi10_40cm_inst (10–40 cm), SoilMoi40_100cm_inst (40–100 cm), and SoilMoi100_200cm_inst (100–200 cm). SWE and canopy water storage were obtained from SWE_inst and CanopInt_inst, respectively. All of these variables are provided in kg m⁻² and were converted to equivalent water thickness before component subtraction. Monthly anomalies were calculated relative to 2004–2009 means to match the GRACE TWSA anomaly baseline. The GLDAS soil-water-storage term used for GWSA inversion is distinct from the separate 1 km surface-soil-moisture dataset used as a high-resolution predictor in Sect. 2.1.4.

To benchmark the reconstructed groundwater-storage anomalies, we selected two ISIMIP global hydrological-model products: h08_20crv3-era5_obsclim_histsoc_default_groundwstor_global_monthly (H08, ISIMIP3a) and watergap2-2c_gswp3-ewembi_nobc_hist_nosoc_co2_groundwstor_global_mon-

thly (WaterGAP 2.2c, ISIMIP2a). Both provide the monthly variable groundwstor in kg m⁻² on a global 0.5° grid. H08 represents natural hydrological processes together with human water use (Hanasaki et al., 2008), whereas WaterGAP 2.2c combines global hydrology with sectoral water-use modules (Müller Schmied et al., 2016). To ensure grid consistency, we bilinearly interpolated both datasets to the CSR GRACE reference grid, retained missing values in the time series, and standardized their analysis periods to April 2002–December 2021 for H08 and April 2002–December 2016 for WaterGAP.

Surface hydrological states and fluxes vary across climatic zones, and GLDAS integrates satellite- and ground-based observations with land-surface modeling to simulate these processes (Rodell et al., 2004). To comprehensively consider various influencing factors, 19 variables closely associated with climate, hydrology, and vegetation dynamics were selected to construct a global environmental variable database at a 1 km scale as modeling predictors. This dataset encompasses 8 monthly-scale dynamic variables, 9 multi-year average variables, and two additional temporal coding variables (derived from sine and cosine transformations of monthly values to capture seasonal characteristics).

Monthly TAS and monthly PRE data are obtained from Climatologies at High resolution for the Earth's Land Surface Areas (CHELSA). This dataset, based on ERA climatic reanalysis and terrain correction algorithms, offers global meteorological forcing data with a spatial resolution of 1 km at the monthly scale, effectively supporting high-precision model simulations at local scales (Karger et al., 2017).

The ET data are sourced from the PML_V2 high-resolution dataset (Zhang et al., 2019), which provides a spatial resolution of 500 m and a temporal resolution of 8 d for ET measurements along with three key independent components (i.e., vegetation transpiration, soil direct evaporation, and canopy interception evaporation). These data have undergone cross-validation in the Global Flux Network (FLUXNET) and exhibit robust representation of surface water consumption patterns. In this study, monthly synthetic ET was generated by rescaling the 8 d resolution data according to temporal scales, followed by cumulative ET calculations to obtain monthly cumulative evapotranspiration (ETC) values since April 2002.

SM was derived from the Global Daily Surface Soil Moisture Dataset (Zheng et al., 2023), which was jointly inverted using ERA5 reanalysis data and ESA-CCI SSM products. This dataset provides seamless global surface soil moisture data with a daily resolution of 1 km for the period 2000–2021. The monthly-scale product was selected.

The DEM is derived from the Global Multi-resolution Topographic Elevation Dataset (GMTED2010) provided by the United States Geological Survey (USGS). This study specifically extracts the spatial mean (Mean) product from this dataset, aiming to precisely characterize the average elevation variation of the underlying surface at the target grid scale.

The soil attributes are derived from the High Precision Global Soil Information System (SoilGrids) published by the International Soil Reference Information Centre (ISRIC). This dataset is generated through machine learning algorithms coupled with multi-source environmental covariate physical coupling inversion, providing global-scale products at a resolution of 250 m. The study extracts variables including Proportion of Clay, Silt, Proportion of Sand, CFVO, CEC, Soil pH, SOC, and TN from depths of 0–30 cm to reflect surface soil (or topsoil) characteristics.

NDVI and LST were derived from MOD13A1 and MOD11A2 products, respectively. This study conducted quality validation based on the Google Earth Engine platform and synthesized monthly average data for the period 2002–2020.

To independently evaluate the spatiotemporal trends in the reconstructed GWSA, we used three types of in-situ groundwater observations: the global groundwater-monitoring network compiled by Jasechko et al. (2024), the Global-Scale Time Series Dataset for Groundwater Studies within the Earth System (GROW; Bäthge et al., 2026), and the China Groundwater Level Yearbook for Geo-environmental Monitoring.

The global monitoring network of Jasechko et al. (2024) comprises 152 069 monitoring wells, with records spanning 1801–2022 and groundwater depth reported at annual intervals. We converted groundwater depth to groundwater-level change by reversing its sign and excluded stations with fewer than eight valid annual observations. Because groundwater-level change and groundwater-storage change have different physical dimensions, an effective specific-yield coefficient was introduced to calibrate the two quantities (Sect. 2.5.1), and its sensitivity was further evaluated (Sect. 3.1.2). In addition, uncalibrated standardized comparisons were provided (Sect. 3.2.1) to assess the consistency between the reconstructed GWSA and global in-situ observations.

The GROW dataset covers 53 countries and contains more than 200 000 groundwater time series, together with 36 Earth-system-related variables and 34 time-series feature flags. Daily and monthly records account for approximately 6 % and 9 % of the dataset, respectively (Bäthge et al., 2026). Following the GROW processing guidelines (Bäthge et al., 2026), we removed yearly records, calculated monthly means from the remaining observations, and retained stations with records throughout April 2002–December 2020. After spatial matching with the WHYMAP global aquifer database and retaining aquifers with more than five valid stations, 53 stations covering three aquifer classes were selected.

The China Groundwater Level Yearbook for Geo-environmental Monitoring compiles daily groundwater-level monitoring records across China. The data cover January 2005–December 2017 and include 1074 monitoring stations with geographic coordinates across 31 provincial-level regions, of which 266 provide continuous observations over the 13-year period. Applying the same screening criteria yielded 87 stations covering aquifers in the North China Plain.

For both datasets, groundwater depth was converted to groundwater-level change, values exceeding three standard deviations were removed, and missing values resulting from the removal were linearly interpolated. Anomalies were then calculated relative to the 2004–2009 and 2005–2009 baseline periods, respectively. The processed station data were subsequently matched to WHYMAP aquifers and a common 1° grid for monthly validation.

To minimize signal leakage errors induced by permanent water bodies and maximize the preservation of coastal terrestrial information, this study employed the Global Surface Water Explorer (GSWE) dataset furnished by the European Commission Joint Research Centre (JRC) (Pekel et al., 2016). This dataset documents the spatiotemporal dynamics of global surface water bodies at a spatial resolution of 30 m. Regions with pixel values of 3 (Permanent water) were extracted from its Yearly Classification product as reliable water body benchmarks. Pixel-wise set union processing was conducted on the temporal classification results spanning from 2002 to 2020, and resampling to a 1 km resolution was carried out, ultimately generating a unified global permanent water body range mask.

The global mountain glacier mask was established by utilizing data from The Randolph Glacier Inventory (RGI) supplied by the National Snow and Ice Data Center (NSIDC) (https://nsidc.org/data/nsidc-0770/versions/7#anchor-documentation, last access: 15 September 2026). This dataset encompasses single-contour vector maps of each glacier dating back to approximately 2000. These maps were integrated within ArcGIS Pro 3.5.4 to demarcate the global mountain glacier extent.

The Greenland land ice extent was obtained from the MEaSUREs Greenland Ice Mapping Project (GIMP) Land Ice and Ocean Classification Mask version 1 data provided by NSIDC (https://nsidc.org/data/nsidc-0714/versions/1, last access: 15 September 2026), making use of the 30 m spatial resolution classification data from the GimpIceMask_30m_2015 product.

Ultimately, the two masks were combined into a unified global glacier mask, which was then merged with the water body mask to generate the baseline land ice extent.

To capture the intricate nonlinear relationships between GWSA and multi-source hydro-climatic variables, two representative machine-learning architectures were employed: XGBoost and DNN. XGBoost is a scalable gradient-boosted tree ensemble that can represent nonlinear statistical mappings (Chen and Guestrin, 2016); machine-learning approaches have also been used to downscale GRACE-derived groundwater signals in regional studies (Miro and Famiglietti, 2018; Seyoum et al., 2019). Its regularization, shrinkage, and row/column subsampling help control model complexity and overfitting (Chen and Guestrin, 2016; Hastie et al., 2009). Complementarily, DNN leverage interconnected hierarchical layers to automatically extract deep-level variable interactions from extensive datasets (LeCun et al., 2015; Shen, 2018). The high model capacity of DNN effectively compensates for the potential limitations of tree-based decision spaces in smoothing nonlinear surfaces, especially under extreme hydrological forcing conditions (Reichstein et al., 2019; Shen, 2018).

To account for global hydrogeological heterogeneity, both models were independently trained for each aquifer category. For each aquifer, 200 000 pixels were randomly sampled for hyperparameter optimization. The model was trained using 70 % of all samples from each aquifer, while the remaining 30 % were used for model accuracy assessment. For XGBoost, the optimal hyperparameters, including n_estimators, learning_rate, max_depth, and min_child_weight, were identified via the RandomizedSearchCV method. For the DNN, a Multi-Layer Perceptron (MLP) architecture was constructed, and a Robust Scaler was applied to both input and target variables to eliminate interference from extreme climate outliers (Shen, 2018). The optimal network configuration (3–6 layers and 128–768 units) was autonomously determined using the Optuna optimization framework (Akiba et al., 2019). The DNN was trained using the Adam optimizer with a mean squared error (MSE) loss function, incorporating an Early Stopping mechanism to prevent over-fitting on the large-scale spatiotemporal dataset.

We evaluated model performance using the coefficient of determination (R2), Pearson's correlation coefficient (CC), bias (Bias), mean absolute error (MAE), and root mean square error (RMSE). R2 and CC quantify the explained variance and linear agreement between predicted and observed values, respectively, whereas Bias, MAE, and RMSE characterize the systematic deviation, mean absolute deviation, and overall magnitude of prediction errors, respectively. 16R2=1-∑i=1n(yi-oi)2∑i=1n(oi-o‾)217CC=∑i=1n(yi-y‾)(oi-o‾)∑i=1n(yi-y‾)2∑i=1n(oi-o‾)218Bias=1n∑i=1n(yi-oi)19MAE=1n∑i=1nyi-oi20RMSE=∑i=1n(oi-yi)2n where o represents the observed value; y represents the predicted value; n represents the sample size; and o‾ and y‾ are the means of o and y respectively.

Monitoring sites with at least eight valid annual observations were retained for the long-term trend evaluation. Groundwater depth was converted to groundwater-level (GWL) change by reversing its sign, such that positive trends indicate groundwater-level recovery and negative trends indicate groundwater-level decline. For each monitoring site, the long-term GWL trend was estimated using the Theil–Sen slope estimator and expressed in cm yr⁻¹. The corresponding long-term GWSA trend was extracted from the 1 km reconstructed product at the same monitoring location.

Because groundwater-level change and groundwater-storage change have different physical dimensions, a scale-dependent effective specific yield, Syeff, was introduced as a level-to-storage conversion coefficient. The purpose of Syeff is to harmonize observation-derived GWL trends with GWSA trends expressed as equivalent water height; it is not intended to represent a direct measurement of local aquifer-specific yield.

To reduce information leakage between coefficient estimation and validation, the monitoring network was partitioned into 1° spatial blocks. Approximately 70 % of the blocks were assigned to the calibration subset and the remaining 30 % to the held-out validation subset. All wells belonging to the same block were retained in the same subset. Syeff was estimated exclusively from observations in the calibration blocks and was subsequently transferred to monitoring sites in the held-out blocks without refitting or using observations from the validation subset.

Within the calibration subset, Syeff was estimated from the multiplicative relationship between long-term GWL trends and the corresponding reconstructed GWSA trends: 21ΔGWSobs=Syeff×ΔGWLobs where ΔGWSobs is the GWS trend derived from the in-situ observations, ΔGWLobs is the long term GWL trend, and Syeff is the effective conversion coefficient used to convert GWL changes into equivalent GWS changes.

The converted in-situ GWS trends and reconstructed GWSA trends were then aggregated separately onto the original 0.25° GWSA grid. Only grid cells containing valid values from both datasets were retained. Anomalous values were removed using a three standard deviation criterion. CC, R2, RMSE, Bias, regression slope, intercept, and trend sign consistency were then calculated. The Syeff coefficient was used only to harmonize the units of the in-situ validation data and did not contribute to the production of the GWSA dataset.

To evaluate uncertainty in the conversion coefficient, the calibration derived Syeff values were perturbed by -30 %, -15 %, +15 %, and +30 %. The complete calculation procedure was repeated for each perturbation scenario. Changes in correlation, error, regression amplitude, and trend sign consistency were used to quantify the influence of Syeff uncertainty on the validation results.

Conversion free in-situ validation was conducted at annual and monthly scales. For the annual scale analysis, the GWL and reconstructed GWSA series for each aquifer were converted to anomalies relative to the 2004 to 2009 reference period, smoothed using a three year moving average to reduce short term interannual fluctuations, and independently standardized. Pearson correlation coefficients were then compared among different aquifer categories and representative aquifer systems to evaluate the consistency of interannual trajectories and long term covariation without introducing uncertainty from the Syeff conversion.

The monthly scale analysis was designed to further evaluate the ability of the reconstructed GWSA to reproduce long term trends, seasonal variations, and short term fluctuations. Monitoring sites with continuous observations throughout the analysis period were first selected. Aquifers containing more than five valid monitoring sites were then retained. Anomalous values in the continuous monthly series were removed using a three standard deviation criterion. The remaining station series were aggregated to the aquifer scale using the median, and the corresponding reconstructed GWSA series were extracted and aggregated over the same aquifer domains.

The aquifer scale observation and reconstructed GWSA series were independently standardized and decomposed using STL into long term trend, seasonal, and residual components. Pearson correlation coefficients were calculated for the original monthly series, trend components, and seasonal components. Seasonal contributions were calculated following the method described in Sect. 2.6. This procedure evaluates the consistency of monthly variability, long term evolution, and seasonal phase, and provides a basis for assessing the ability of the reconstruction to represent regional signals and local seasonal processes across different aquifer categories.

The 1 km GWSA product was compared with published estimates for the North China Plain, northwestern India, California, and the High Plains Aquifer. The spatial extent of each region was reconstructed as closely as possible to the corresponding reference study. Public numerical data were used preferentially, and published curves were digitized only when numerical data were unavailable. The reconstructed and reference series were matched in time and adjusted to a common baseline according to the temporal coverage and data structure of each study.

CC, RMSE, Bias, standard deviation ratio, seasonal amplitude, and trend slope were used to evaluate regional temporal variability, amplitude agreement, and long term trend consistency. For studies reporting a specific analysis period, ordinary least squares trends were additionally calculated over the same time window. This regional comparison links the global 1 km reconstruction with well studied groundwater systems, including the North China Plain, northwestern India, California, and the High Plains Aquifer. It provides a regional assessment of the product's ability to represent persistent groundwater depletion, seasonal variability, and trend magnitude, and establishes a common basis for comparison with existing regional knowledge.

To ensure the spatiotemporal continuity of the downscaling baseline without resorting to artificial annual averaging, SSA was employed to reconstruct missing GRACE observations (Fig. 2). Global evaluation confirms that the SSA framework successfully maintained low reconstruction errors while preserving the integrity of the original temporal dynamics across two distinct gap-filling phases.

During the SSA-filling-a phase, the reconstruction error was quantified as the standard deviation of residuals between observed and reconstructed values over all non-missing months. Globally, the error ranged from 0.232 to 62.458 cm, with a tightly constrained median of 1.156 cm. The 1 % and 99 % percentiles were 0.475 and 7.550 cm, respectively, indicating robust reconstruction precision during this stage.

To empirically represent the reconstruction capability under conditions of long-term continuous missing data, a rigorous cross-validation approach was implemented in the second phase. Specifically, entire years of data within known observation periods were sequentially masked and reconstructed. At the global scale, the median error for this phase was 1.679 cm, with the 1st and 99th percentiles at 0.732 and 10.204 cm, respectively. Focusing specifically on terrestrial regions (excluding Antarctica), the median error was 2.708 cm, with 1st and 99th percentiles of 0.862 and 15.295 cm. While the errors in land areas are slightly higher than the global average, they remain consistently low and bounded.

Spatially, the high-error centers across both phases exhibited high consistency, primarily clustering in the high-latitude cryosphere of the Northern Hemisphere, the Amazon Basin, southern South America, East Africa, and Southeast Asia. Although uncertainties in these specific regions are relatively higher, these variations are primarily associated with large natural signal amplitudes, significant seasonal oscillations, and complex local processes rather than algorithmic instability. Consequently, the SSA reconstruction provides a temporally continuous baseline with quantified reconstruction uncertainty for the subsequent high-resolution downscaling.

Table 2 first compares the two candidate architectures within the random 70 %/30 % development split for each WHYMAP mega-aquifer system or specific mask. XGBoost yielded lower RMSE and MAE and higher R2 than DNN in all 11 listed aquifer systems; for AQU 11, for example, XGBoost achieved RMSE = 1.15 cm and R2=0.921, compared with 1.61 cm and 0.845 for DNN. These values are within-distribution diagnostics used for model comparison, while product reliability is evaluated below through cross-scale consistency and spatially held-out in-situ observations.

Cross-scale aggregation to the original 0.25° grid confirmed the same overall ranking (Fig. 3). At the global scale, XGBoost achieved R2 = 0.972, CC = 0.986, and RMSE = 2.097 cm, compared with R2 = 0.961, CC = 0.981, and RMSE = 2.661 cm for DNN. Across the 11 evaluated aquifer categories, XGBoost achieved higher R2 and CC and lower RMSE in 10 categories. Major groundwater basins (AQU categories 11–15) and regions with complex hydrogeological structures (AQU categories 22–25) maintained an average R2 of 0.961 and RMSE of 0.730 cm, whereas local and shallow aquifers (AQU categories 33–34) had a higher average RMSE of 2.091 cm. The cross-scale assessment therefore shows that the relative performance advantage of XGBoost is retained when the 1 km reconstructions are evaluated against the original-resolution GWSA.

A global observation-constrained in-situ evaluation was conducted using long-term groundwater-monitoring records. Following the 1° spatial-block design described in Sect. 2.5.1, the comparison was performed on retained 0.25° validation grids after dimensional harmonization, common-grid aggregation, and uniform quality control. Figure 4 summarizes the resulting grid-level agreement for the DNN and XGBoost products.

For DNN, CC was 0.72, corresponding to R2 = 0.52, with RMSE = 0.34 cm yr⁻¹ and Bias = 0.09 cm yr⁻¹. The fitted regression slope and intercept were 0.51 and -0.04 cm yr⁻¹, respectively, and trend-sign consistency was 0.72. XGBoost exhibited slightly stronger spatial covariance and directional consistency, with CC = 0.73, R2=0.54, RMSE = 0.38 cm yr⁻¹, Bias = 0.11 cm yr⁻¹, a fitted slope of 0.51, an intercept of -0.05 cm yr⁻¹, and trend-sign consistency of 0.73. Although slopes below unity indicate residual differences in the absolute magnitude of long-term groundwater-storage change, both models reproduced the dominant spatial covariance and direction of groundwater depletion and recovery across the monitoring network. XGBoost showed slightly stronger correlation and trend-direction agreement, whereas DNN exhibited marginally lower absolute-error statistics. Together with the cross-scale consistency assessment, these results support the use of XGBoost as the production architecture without implying uniform superiority across every metric.

The in-situ evaluation complements the cross-scale consistency assessment by providing an external comparison with groundwater-monitoring records. The 0.25° grid-based implementation reduces the mismatch between point-scale well observations and spatially integrated GWSA estimates, while the Z-score comparisons without Syeff in Figs. 6–8 provide an additional assessment of temporal co-variation.

The sensitivity of the in-situ evaluation to Syeff was further quantified by perturbing the baseline conversion coefficients by ± 15 % and ± 30 % (Table 3). For the XGBoost product, the baseline evaluation yielded CC = 0.734, R2 = 0.538, RMSE = 0.378 cm yr⁻¹, Bias = 0.106 cm yr⁻¹, and a regression slope of 0.506. Across the full ± 30 % perturbation range, CC varied from 0.691 to 0.734 and R2 from 0.477 to 0.538, whereas RMSE remained within 0.359–0.408 cm yr⁻¹ and Bias within 0.104–0.107 cm yr⁻¹. The regression slope showed a comparatively stronger response, ranging from 0.379 to 0.593, while trend-sign consistency remained between 0.696 and 0.743. DNN exhibited a similar response, with R2 ranging from 0.476 to 0.522 and RMSE from 0.343 to 0.393 cm yr⁻¹. These results indicate that uncertainty in Syeff has a stronger influence on amplitude-related quantities, particularly the regression slope, than on the overall association and trend-direction agreement between the modeled and in-situ-derived groundwater-storage trends.

Taken together, the aquifer-stratified internal diagnostics, cross-scale consistency assessment, and spatially independent in-situ evaluation support XGBoost as the production architecture. XGBoost provided more consistent performance across aquifer categories and stronger global cross-scale agreement, while DNN retained slightly lower absolute errors in the in-situ trend comparison. The selection therefore reflects the combined evidence rather than uniform superiority on every metric.

The target-product experiments show that model fit was reproducible but depended on the source product (Fig. 5). Across the 12 aquifer categories and masks, mean internal-test R2 was 0.928 for CSR, 0.946 for JPL, 0.888 for WaterGAP, and 0.908 for H08. JPL produced the highest R2 in 10 categories, while CSR and WaterGAP each ranked first in one category. The full ranges were 0.900–0.994 for CSR, 0.915–0.996 for JPL, 0.794–0.924 for WaterGAP, and 0.828–0.941 for H08. All four products produced high within-product fit, with systematic differences among target sources.

Predictor importance also depended on the target source. Static physiography and soil variables accounted for 0.638 and 0.620 of total importance in the CSR and JPL experiments, compared with 0.454 and 0.442 for WaterGAP and H08. Conversely, temporal encoding contributed 0.102 and 0.107 for CSR and JPL but increased to 0.191 and 0.225 for WaterGAP and H08. These differences indicate that the relative importance assigned to static heterogeneity and temporal variation depends on the target source and represents a source-related structural uncertainty.

To assess temporal fidelity under diverse hydrogeological and anthropogenic settings, complementary in-situ comparisons were categorized into climate-driven, anthropogenic-dominated, and data-limited basins (Figs. 6–8). These analyses compare independently standardized GWL and downscaled GWSA time series, and thus provide a conversion-free assessment of interannual variability and long-term temporal co-variation.

In natural, climate-dominated regions (Fig. 6), the downscaled dataset demonstrates exceptional temporal alignment with ground observations. For instance, within the Stampriet-Kalahari Basin (AQU 12, N = 7849), the localized time-series comparison yielded a high CC = 0.82, accurately reproducing the multi-year depletion trajectory from 2004 to 2014 and subsequent recovery (Fig. 6b). Similarly, robust agreement was observed in the Cuvelai-Upper Zambezi Basin (AQU 11, N = 1713, CC = 0.72, Fig. 6c), the Russian Platform Basins (AQU 33, N = 32340, CC = 0.67, Fig. 6f), and the Ganges-Brahmaputra Basin (AQU 24, N = 2208, CC = 0.54, Fig. 6d). Although the Arabian Aquifer System (AQU 22, N = 1377, Fig. 6e) exhibited a weaker linear correlation (CC = 0.07) due to complex local fossil groundwater dynamics, the overall multi-year deficit trend remains observable. These findings confirm that the machine learning framework effectively translates macro-scale climatic forcing into high-resolution, localized groundwater storage variations.

Conversely, evaluations in intensive anthropogenic basins (Fig. 7) reveal distinct dynamic behaviors. Notably, in the Indus Basin (AQU 23, N = 1383, Fig. 7c), the application of the localized mask and variance-normalized Z-scores demonstrates a highly consistent temporal trajectory (CC = 0.58), successfully capturing the severe depletion trough around 2013 and the subsequent recovery phase driven by regional water management. Meanwhile, in the Northern Great Plains Aquifer (AQU 14, N = 10052, CC = -0.23, Fig. 7b) and the North Caucasus Basin (AQU 34, N = 3971, CC = -0.01, Fig. 7e), predicted climate-driven baselines indicate storage recovery or dynamic stability responding to favorable meteorological conditions, whereas in-situ observations exhibit sustained depletion. This decoupling indicates that the downscaled GWSA and local monitoring wells may respond differently in intensively managed aquifers. The resulting discrepancy is therefore interpreted as a signal potentially consistent with anthropogenic groundwater depletion rather than as a direct estimate of groundwater extraction.

In data-limited or complex regions (Fig. 8), such as the Cambrian-Ordovician Aquifer (AQU 15, N = 634, CC = 0.24, Fig. 8b) and the West Siberian Basin (AQU 25, N = 9, CC = 0.48, Fig. 8c), the validation metrics reflect the inherent challenges of sparse observation networks and complex hydrogeological constraints, yet the downscaled product still manages to capture the broad multi-year wave patterns corresponding to the limited in-situ records.

Figure 9 provides a monthly-scale validation of the reconstructed GWSA using five representative aquifer-site groups distributed across North America, Europe, and East Asia. The maps show the aquifer masks and the available monitoring sites, while the time-series panels decompose the standardized station and reconstructed records into STL trend, seasonal, and residual components. Four records span April 2002–December 2020; the ID 1417 record spans 2005–2017 because of its available observations.

To further quantify the monthly-scale consistency between the reconstructed GWSA and in-situ observations, we first aggregated the monitoring-site records to the aquifer scale using the median method. The GWSA and aquifer series were then standardized to remove dimensional differences. Finally, STL decomposition was applied to separate the original series into long-term trend and seasonal components, and correlation coefficients for the original, trend, and seasonal series, together with the seasonal contribution, were calculated.

Overall, the mean correlation coefficients across the five groups were 0.51, 0.55, and 0.39 for the original, trend, and seasonal series, respectively. This indicates that the reconstruction reproduces long-term variability slightly better than monthly and seasonal fluctuations. The highest mean correlations were obtained for major groundwater basins (AQU 13; 0.75, 0.85, and 0.13), followed by aquifers with complex hydrogeological structures (AQU 24; 0.41, 0.57, and 0.46). Local and shallow aquifers (AQU 34) showed the lowest correlations for the original and trend series (0.33, 0.27, and 0.61). These differences indicate that coherent basin-scale storage variations favor agreement with in-situ observations, whereas local hydrogeological heterogeneity, strong seasonality, and anthropogenic perturbations reduce temporal consistency.

Major groundwater basins (AQU 13) include the High Plains Aquifer (ID 1308) and the North China Plain (ID 1417), with correlation coefficients of (0.67, 0.76, and 0.00) and (0.82, 0.91, and 0.26), respectively. In the High Plains Aquifer, the reconstruction captures the transition from early depletion to later recovery, and its seasonal contribution is close to that of the in-situ observations (0.27 and 0.23). However, the near-zero seasonal correlation suggests a phase offset or lag in the seasonal cycle. In the North China Plain, the reconstruction reproduces the persistent groundwater deficit and shows a similar seasonal contribution to the in-situ observations (0.22 and 0.26), probably because the large alluvial plain produces a relatively coherent regional response despite intensive irrigation and groundwater abstraction. Aquifers with complex hydrogeological structures (AQU 24) are represented by the Floridan Aquifer (ID 1669), for which the correlations are (0.41, 0.57, and 0.46). The reconstructed and in-situ seasonal contributions are relatively close (0.40 and 0.47), whereas complex carbonate-aquifer architecture and spatially variable recharge may weaken the correspondence between local observations and the regional reconstruction.

Local and shallow aquifers (AQU 34) include the Apennine Peninsula aquifer systems (ID 1193) and the Scandinavian Peninsula aquifer systems (ID 426), with correlation coefficients of (0.36, 0.02, and 0.70) and (0.29, 0.51, and 0.51), respectively. In the Apennine region, the reconstruction retains substantial seasonal co-variation but fails to reproduce the long-term trend. This may reflect strong local recharge contrasts associated with mountainous terrain, carbonate and fractured aquifers, together with local pumping effects. In the Scandinavian region, the reconstruction partly captures the overall trend, but its monthly and seasonal agreement is weaker, probably because of pronounced snowmelt and freeze–thaw cycles, shallow groundwater responses, sparse monitoring, and local hydrogeological heterogeneity. Thus, the low original-series correlations in both AQU 34 regions mainly reflect the influence of local processes and high-frequency seasonal variability on monthly matching.

The seasonal-contribution results show that the mean contribution is approximately 0.39 for the reconstruction, lower than approximately 0.48 for the in-situ observations. In AQU 13, the reconstructed and in-situ seasonal contributions are relatively close for both the High Plains Aquifer and the North China Plain. The difference is also small in AQU 24. By contrast, in AQU 34, the in-situ contributions reach 0.74 and 0.72 in the Apennine and Scandinavian regions, respectively, compared with 0.49 and 0.58 for the reconstructions. The product therefore captures the regional seasonal rhythm of large basins more effectively but tends to underestimate seasonal amplitudes in local and shallow aquifers, particularly in high-latitude environments.

Figure 10 and Table 4 compare the unsmoothed regional series on matched dates. Agreement was strongest in the NCP (r = 0.963; RMSE = 32.0 mm), followed by California (r = 0.789; RMSE = 77.3 mm) and northwestern India (r = 0.731; RMSE = 66.4 mm). The NCP had the closest total variability magnitude (SD ratio = 0.912), whereas variability was attenuated in northwestern India (0.593) and California (0.673). Product and reference seasonal amplitudes were similar in California (82.0 and 82.7 mm), but the product had stronger seasonality in the NCP (23.2 versus 8.9 mm) and northwestern India (63.9 versus 23.5 mm). Thus, strong correlation did not guarantee agreement in seasonal or total amplitude.

The paired-series Sen slopes had the same negative sign in the three monthly comparisons, but depletion was weaker in the product: -17.71 versus -20.27 mm yr⁻¹ in the NCP, -22.96 versus -39.47 mm yr⁻¹ in northwestern India, and -6.96 versus -15.80 mm yr⁻¹ in California. Exact-window OLS comparisons likewise gave -13.74 mm yr⁻¹ for NCP during 2003–2016, versus -16.1 mm yr⁻¹ reported by Zhang et al. (2023); -24.47 mm yr⁻¹ for northwestern India during August 2002–October 2008, versus -40.0 ± 10.0 mm yr⁻¹ in Rodell et al. (2009); and -6.93 mm yr⁻¹ for California during October 2003–March 2010, versus -20.4 ± 3.9 mm yr⁻¹ in Famiglietti et al. (2011). The full 2003–2021 trend of Liu et al. (2022) was not treated as an exact-window test because the product ends in December 2020.

The HPA comparison was less favorable. The annual well-based composite from Rateb et al. (2020) yielded r = 0.526 and RMSE = 48.2 mm. The product Sen slope was nearly neutral (-0.10 mm yr⁻¹; 95 % confidence interval -5.10 to +5.41 mm yr⁻¹), compared with -9.28 mm yr⁻¹ in the reference. More importantly, over the exact March 2003–February 2013 window of Breña-Naranjo et al. (2014), the product OLS trend was +4.50 mm yr⁻¹, opposite to the published -27.6 ± 3.3 mm yr⁻¹. The reconstruction therefore did not reproduce the reported HPA-wide depletion during that interval. These contrasts identify regional limitations in depletion magnitude and show why literature benchmarking is informative beyond correlation-only evaluation.

The spatial pattern of the global Mean GWSA was reconstructed at a 1 km resolution for the period from 2002 to 2020, along with the corresponding meridional mean profiles (Fig. 12). This reconstruction clarifies the baseline state of mass balance in the global groundwater system during the study period and offers deeper insights into the current imbalanced distribution patterns of the terrestrial water cycle on Earth. From the perspective of the global spatial pattern (Fig. 12b), the TWS exhibits systematic mass imbalances, manifested primarily as a net surplus in tropical and mid-to-high-latitude humid zones, whereas cumulative depletion occurs in areas of intensive human intervention and arid regions.

Significant negative anomalies are not randomly distributed but are highly concentrated in several typical large alluvial plain aquifers and arid regions globally. These localized depletion centers include the Indus-Ganges Basin, the North China Plain, the Arabian Peninsula, the Caspian Sea region, and the High Plains of the United States. In sharp contrast, significant positive anomalies, functioning as global groundwater accumulation zones, are stably distributed in tropical rainforest areas influenced by the Intertropical Convergence Zone (ITCZ) and certain mid-to-high-latitude regions affected by persistent moist air currents.

The meridional mean profile (Fig. 12a) further condenses the two-dimensional heterogeneity into distinct zonal distribution characteristics along the latitudinal gradient. A core feature of this profile is the global groundwater depletion zone situated between 25 and 45° N, with its trough spatially corresponding to groundwater depletion hotspots located in regions characterized by intensive human water use on the map. From the perspective of the dynamic mechanism, this pattern is spatially associated with the combined influence of climatic drought stress and intensive human water use. This latitude belt not only lies within the long-term recharge-deficient region associated with the descending branch of the Hadley Cell but also represents the most densely populated and agriculturally irrigated area globally.

The reduction in the depth of the mean profile indicates that the groundwater system in this zone has fundamentally deviated from its climate-driven natural equilibrium state, leading to persistent mass loss. The profile exhibits stable positive peaks near the equator, consistent with the high concentration of reserves in tropical rainforest regions, while a pronounced negative shift occurs at extremely high latitudes (around 80° N). This shift effectively captures land mass loss signals associated with cryosphere retreat or related dynamic processes at the polar margins. Ultimately, this spatial distribution of multi-year averages serves as a physical framework for understanding the evolution of the global groundwater system, providing essential contextual support for the subsequent analysis of dynamic evolution rates.

To further evaluate the influence of cyclic factors on the dynamics of global groundwater storage, this study quantified the relative contributions of seasonal and non-seasonal components using the downscaled monthly GWSA time series (Fig. 13). In this framework, seasonality is characterized by the average intra-annual standard deviation for each year from 2002 to 2020, whereas non-seasonality is measured by the standard deviation of the annual mean values over the same 19-year period. The relative seasonal contribution is determined by the ratio of the seasonal component to the sum of both components, with a threshold of 50 % used to identify regions predominantly governed by seasonal factors.

The results indicate that global GWSA dynamics generally exhibit distinct seasonally-regulated characteristics. Based on equal-area projection statistics, the seasonal contribution exceeds 50 % across 52.30 % of the global land area. This suggests that over half of the terrestrial domains are predominantly governed by stable and recurrent intra-annual cycles, rather than non-seasonal anomalies at interannual scales. Furthermore, in 26.09 % of the land area, the seasonal contribution is greater than 66 %, meaning the seasonal component is more than 1.5 times the magnitude of non-seasonal fluctuations, indicating a pronounced dominance of seasonality.

Spatially, regions with high seasonal contributions (60 % to 90 %) are primarily concentrated in areas with regular hydrological rhythms, including the Amazon and Congo Basins, the South and East Asian monsoon regions, northern Europe to the southern margin of Siberia, and high-latitude zones in North America. These patterns reflect a robust response of groundwater storage to periodic variations in precipitation, evapotranspiration, surface water-groundwater interactions, and freeze-thaw processes.

In contrast, regions with low seasonal contributions are predominantly distributed in inland arid zones, such as North Africa, the Arabian Peninsula, Central Asia, inland Australia, and western North America. In these areas, non-seasonal fluctuations dominate, manifesting primarily as persistent GWSA deficits associated with long-term trends. The frequent occurrence of extreme heat, compound droughts, and heatwave events in these regions exacerbates long-term soil moisture imbalances. These groundwater systems are heavily associated with persistent drought, intensive overexploitation, and complex topography-climate coupling. Consequently, they have transitioned from structured, seasonal recharge-depletion cycles to a state dominated by non-periodic disturbances.

As illustrated in Fig. 14, the evolution of global groundwater storage over the past two decades has exhibited notable regional clustering characteristics and spatial differentiation trends. With a spatial resolution of 1 km, the downscaled trend maps reduce the visual effects of spatial smoothing inherent in traditional coarse-resolution GRACE observations and reveal predictor-informed local hydrological gradients under coarse-scale GRACE constraints.

Globally, significant groundwater depletion signals (p<0.01 and trend rate < -0.10 cm month⁻¹) are mainly manifested in two typical spatial patterns.

The first pattern consists of relatively clustered, patchy, and linear depletion centers, the spatial distribution of which shows a high degree of overlap with global high-intensity agricultural irrigation basins. For example, the 1 km dataset effectively captures steep depletion gradients along specific pre-mountain alluvial fans or typical farmland boundaries in regions such as the Central Valley of California, the High Plains Aquifer (Ogallala), northwestern India, and the North China Plain of China. This spatial characterization does not establish accurate basin-wide depletion magnitude: the separate regional benchmark in Sect. 3.2.3 identifies an HPA-wide trend mismatch over 2003–2013.

The second pattern involves large-scale, contiguous regional depletion, primarily concentrated in arid and semi-arid climate zones and major fossil groundwater extraction areas, such as the Middle East, the Arabian Peninsula, the hinterland around the Caspian Sea, and the La Plata Plain at the border of Argentina and Brazil in South America. Most of the Middle East region is characterized by a tropical desert climate, while the hinterland around the Caspian Sea is located deep within the continent, featuring arid conditions and intense evaporation, with inherently scarce water resources. The dense distribution of significant depletion points in these areas (indicated by black stippling in Fig. 14) indicates that their groundwater systems remained in a state of sustained and significant depletion during the study period.

Spatially corresponding to the aforementioned deficit patterns, significant groundwater accumulation regions worldwide (p<0.01 with trend rates >0.10 cm month⁻¹) predominantly display diffuse spatial distribution characteristics in accordance with major surface water systems and natural topography. In equatorial and tropical monsoon climate zones, including the transverse extension areas of Africa's Sahel region, the Victoria Lake Basin, and the core Amazon Basin, the GWSA demonstrates extensive and spatially continuous significant surplus. Additionally, high-latitude regions in North America, such as northern Canada with a polar long-winter climate, and localized areas in northern Eurasia, such as the southeastern Siberian Plain in Russia where permafrost is prevalent, show groundwater accumulation signals in trend maps. These signals potentially reflect deep water infiltration associated with accelerated hydrological cycles or cryosphere degradation at high latitudes. These surplus zones, dominated by large-scale climatic or natural hydrological processes, exhibit relatively smooth spatial transitions at 1 km scales, forming distinct morphological contrasts with the sharper deficit boundaries spatially coincident with intensively managed agricultural regions.

To summarize the magnitude and statistical persistence of global groundwater trends, Theil–Sen trend rates were combined with MK significance thresholds to define seven descriptive classes (Fig. 15). This classification identifies the direction, rate, and significance category of each mapped trend; it does not by itself diagnose the external forcing responsible for that trend.

Pixel-level statistical results indicate that groundwater reserves in most global regions have significantly deviated from long-term dynamic equilibrium states from 2002 to 2020. Data reveal that approximately 38.84 % of the effective land area is experiencing significant or extremely significant groundwater depletion (p<0.05 or p<0.01), whereas about 49.16 % of regions show significant or extremely significant recharge. In contrast, only 12.00 % of land areas maintained a relatively stable dynamic equilibrium with non-significant trends during the study period.

The MK-based classification indicates that statistically significant positive and negative trends occupy large fractions of the mapped land area (Fig. 15). Among areas classified as depleted, 37.13 % of the total effective land area falls in the p<0.01 category and 1.70 % in the p<0.05 category; the corresponding accumulation fractions are 47.01 % and 2.15 %. These proportions describe the adopted significance classes and should not be interpreted as proof of abrupt physical thresholds or a unique forcing mechanism.

The mapped contrast between non-significant and highly significant classes may reflect a combination of persistent regional trends, the discrete MK thresholds used for classification, predictor-informed spatial structure, and residual-correction effects. Accordingly, sharp class boundaries in the 1 km product are treated as statistical features requiring independent verification, not as direct evidence of abrupt aquifer boundaries, structural overextraction, or any single causal process.