Integrating remote sensing and hybrid machine learning for developing a gridded groundwater storage dataset for different climate scenarios in Sri Lanka

Abstract

Reliable groundwater monitoring is essential to ensure water security. Yet, many regions lack dense observational networks. High-resolution gridded groundwater datasets therefore provide valuable alternatives for characterising groundwater dynamics. The Gravity Recovery and Climate Experiment (GRACE) offers large-scale terrestrial water storage anomaly (TWSA) data, but its coarse spatial resolution limits regional applicability. Recent studies have increasingly utilized machine learning (ML) to improve GRACE-based groundwater estimates; however, most have relied on individual ML models, which enhance prediction accuracy only to a limited extent. Based on these grounds, this study proposes a stacking ensemble machine learning (SEML) framework that integrates Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Categorical Boosting (CatBoost) for more precise statistical downscaling of GRACE-derived groundwater storage anomalies (GWSA) to finer spatial resolutions. The framework was first evaluated in the Kumbukkan Oya Basin, where GWSA was downscaled from 0.25° to 0.05° resolution, achieving strong predictive skill (R² = 0.84, RMSE = 3.04 cm), with grid-wise validation against GRACE yielding R² values exceeding 0.9 and in-situ groundwater level comparisons showing correlations above 0.7 in most wells. To further assess the physical consistency of the downscaled signals, an independent Physics-Informed Neural Network (PINN) framework was developed for the same basin. The groundwater system was idealised as a one-dimensional transient unconfined flow, reflecting the predominantly linear basin geometry. The PINN model embedded the Boussinesq groundwater flow equation directly into the learning process and was trained using rainfall as the primary forcing, with the uppermost and lowermost wells prescribed as transient boundary conditions and two interior wells used for training and testing. The model reproduced historical groundwater dynamics with a correlation coefficient of 0.81 and an RMSE of 0.57 m at the testing interior well, providing physics-based validation of the downscaled GWSA patterns. The methodology was subsequently extended to the entire Sri Lankan domain, generating the first high-resolution national GWSA dataset capable of resolving spatial heterogeneity and groundwater stress patterns. Future groundwater dynamics were assessed using precipitation projections from the CMIP6 CNRM-CM6-1 model under SSP2-4.5 and SSP5-8.5 scenarios for near- (2021–2039) and mid-century (2040–2059) periods. Trend analysis revealed spatially coherent groundwater depletion hotspots, particularly across the dry and intermediate zones, with projected GWSA declines locally exceeding −1.5 cm yr⁻¹ under high-emission conditions. By integrating ensemble learning, climate projections, and national-scale analysis, this study delivers a robust, transferable framework for groundwater monitoring, climate impact assessment, and evidence-based groundwater management in data-limited regions.