Spatio-temporal analysis of groundwater from machine learning based GRACE downscaling in the Moroccan context

Abstract

Morocco is among nations facing high to extremely high-water stress in the world under climate change whose impacts include reduced precipitation, higher temperatures, and evapotranspiration. This pressure is exacerbated by changes in land use and cover, such as agricultural extensification and urbanization, which reduces recharge and increases water demand. However, effective management of water resources, particularly groundwater, remains difficult because of the scarcity of in situ measurements and the low quality of the available data. In this regard, our study aims to downscale GRACE satellite-derived Total Water Storage Anomalies (TWSA) from their native coarse resolution (3°) to a finer resolution (0.04°), followed by Groundwater Storage Anomalies (GWSA) extraction. This is achieved by using machine learning models and hydrological variables that are strongly correlated with TWSA, including precipitation, evapotranspiration, soil moisture, and runoff. The performance of four machine learning models in capturing the spatial details of TWSA is then evaluated, namely, Convolutional Neural Network (CNN), Random Forest (RF), XGBoost, and Gated Recurrent Unit (GRU). Then the results generated by the model with the best results are analyzed spatially and temporally to better understand the trends, the availability, and the influence of LULC changes on groundwater resources. RF delivered the best performance by achieving an R² of 0.92 and an RMSE of 0.61 cm. The RF based TWSA estimates showed strong agreement with ground measurements across different spatial and temporal scales. And the analyses highlight the importance of integrating hydrological and land use factors into groundwater modeling and demonstrate that machine-learning-based downscaling can effectively capture groundwater variability and help bridge the gap between satellite observations and local scale sustainable water management.