SAINT-REG: a deep learning framework for enhancing GRACE data resolution to predict groundwater storage variation with low-volume constraint data

Groundwater is a vital resource supporting agriculture, drinking water, and ecological balance especially in water-stressed regions such as India. However, increasing stress from over-extraction, erratic rainfall, and land-use change has led to alarming groundwater depletion. The National Aeronautics and Space Administration’s Gravity Recovery and Climate Experiment (GRACE) satellite mission offers essential data on terrestrial water storage, yet its coarse spatial resolution (300 km) and limited labeled samples constrain its application for regional-scale groundwater analysis. “Small data problem” is addressed in remote sensing area by downscaling GRACE-derived liquid water equivalent (LwE) thickness values (300-km resolution) using high-resolution auxiliary hydrological variables, including soil moisture from the root zone and surface layers Global Land Data Assimilation System (GLDAS, 25-km resolution), precipitation Climate Hazards Group Infrared Precipitation with Stations (CHIRPS, 5-km resolution), and historical GRACE LwE data. These were extracted from CHIRPS and GLDAS satellite products (in.tif format) and transformed into vector format for modeling. The Self-Attention Inspired Network Transformer for Resolution-Enhanced GRACE Prediction (SAINT-REG) framework was proposed, a deep learning model designed to capture complex spatial and temporal dependencies through dual-attention mechanisms, enabling high-resolution groundwater prediction from coarse GRACE data. The model was trained on 6 years of data (2018 to 2023) across multiple spatial grid points. The improvement in prediction accuracy is reflected in the reduced mean absolute error (MAE = 1.45) of the prediction error, demonstrating that SAINT produces predictions closer to the actual observations than conventional approaches. We highlight the potential of attention-based architectures to overcome data sparsity and support localized groundwater forecasting and resource management.