Machine-learning-based regional-scale groundwater level prediction using GRACE

The rapid decline of groundwater levels (GWL) due to pervasive abstraction in the densely populated (~1 billion) Indus-Ganges-Brahmaputra-Meghna (IGBM) transboundary river basins South Asia, necessitates a robust framework prediction and understanding. While few localized studies exist, three-dimensional regional-scale characterization GWL is yet be implemented. Here, ‘support vector machine’, machine-learning-based method, applied data from Gravity Recovery Climate Experiment (GRACE) on land-surface-model-based storage meteorological variables, predict anomaly (GWLA) IGBM. study has three main objectives, (1) understand spatial (observation well locations) subsurface (shallow vs. deep observation wells) variability results for in-situ GWLA large number wells (n = 4,791); (2) determine its relationship with abstraction, and; (3) outline advantages limitations using GRACE predicting GWLAs. findings, based individual results, suggest significant efficiency (median statistics: r > 0.71, NSE 0.70; p < 0.05) most IGBM; however, identifies hotspots, mostly agriculture-intensive regions, having relatively poor model performance. Further analysis depth-wise statistics reveals that dominance pumping deeper depths aquifer linked performance (screen depth 35 m) compared shallow m), thus, highlighting limitation representing depth-dependent local-scale pumping.