Assessment of high-resolution groundwater storage anomaly using machine learning models in the state of Jharkhand

ABSTRACT Groundwater resources in Jharkhand, India, are under increasing stress due to escalating demand from urban, peri-urban, and agricultural sectors, requiring improved spatial characterisation of groundwater dynamics. GRACE and GRACE-FO provide observations of terrestrial water storage anomaly (TWSA); their coarse grid distribution (∼25 km) limits district-level assessments. This study develops a machine learning–based framework to generate GRACE-derived TWSA at a ∼4 km grid spacing across Jharkhand for 2003–2023 using spatially detailed TerraClimate hydroclimatic predictors. Multiple linear regression, artificial neural network, random forest (RF), extreme gradient boosting, long short-term memory, and support vector regression. RF performed well, achieving R2 and Nash–Sutcliffe efficiency of 0.86 in leave-one-district-out validation. In temporal validation, RF achieved R2 = 0.85 during training (2003–2014) and 0.68 during testing (2015–2023), with root mean square error of 6.23 cm and 9.30 cm, respectively. Groundwater storage anomaly (GWSA) was estimated by subtracting soil moisture anomaly from the downscaled TWSA. Independent in situ validation with model GWSA showed district-level R2 values of 0.52–0.76. The resulting downscaled GWSA revealed that predominantly negative trends occurred in the northeastern districts of Jharkhand. The mean trend was −0.093 cm yr−1, with a 95% Monte Carlo interval of −0.118 to −0.066 cm yr−1.