Hybrid hydroinformatic assessment of groundwater-storage depletion in Saudi Arabia (2002–2024): integrating GRACE/GRACE-FO gravimetry, land-surface modeling, and dependence-aware machine learning

Introduction Saudi Arabia meets most of its water demand from largely non-renewable aquifers, yet the magnitude and spatial organization of storage depletion and the expression of interannual hydroclimate in the storage record have rarely been examined within one baseline-benchmarked framework. Methods We integrate 22 valid years (2002–2024) of GRACE/GRACE-FO mascon gravimetry with GLDAS-2.1 surface storage; CHIRPS precipitation; and MODIS vegetation, temperature, evapotranspiration, and land-cover observations on a 0.5° grid and decompose each cell’s groundwater-storage (GWS) anomaly into a Theil–Sen trend and a detrended residual. Results Kingdom-mean GWS declined at –0.91 cm yr -1 (95% CI –0.93 to –0.88; –19.8 cm in total, ≈4.3 × 10 2 km 3 , range 2.5−4.3 × 10 2 km 3 across surface-storage treatments), with a median detrended residual variance fraction of 4.3%. Under leave-one-year-out validation, five machine-learning models do not exceed a climatological baseline for the residual (best R 2 = –0.02; 0.05 with fold-wise detrending), whose magnitude (≈1.5 cm RMSE) is comparable to GRACE annual-mean uncertainty. A rate-based Pettitt test detects no Kingdom-wide change point in the depletion rate (p = 0.96). Regional interrupted-time-series and difference-in-differences tests regarding the 2016–2018 agricultural reforms show a post-2018 deceleration over the north-western wheat belt (0.42 cm yr -1 , 95% CI 0.09−0.73) that is distinguishable from zero, whereas its contrast against the other agricultural regions (0.31 cm yr -1 , 95% CI –0.07 to 0.70) is not; the pattern is consistent with, but does not demonstrate, a redistribution of depletion. A cross-sectional model reproduces the spatial pattern of depletion rates under leave-spatial-block-out validation (R 2 = 0.60 with cell coordinates; 0.24, 95% CI –0.01 to 0.39, with surface covariates alone); the evapotranspiration trend carries the largest permutation importance, followed by the precipitation climatology and daytime land-surface temperature, with the irrigated-area covariates below them. Discussion The results provide an uncertainty-qualified evidence base for groundwater governance under Vision 2030 and a transferable protocol for hyper-arid, abstraction-dominated regions.