Machine learning algorithms for estimating basin-scale groundwater levels based on GRACE satellite data

Résumé fourni par la source

Study region The study focuses on the Lower Valley of the Medjerda (LVM) Basin in northern Tunisia. Study focus This study aims to predict groundwater levels (GWL) in data-scarce environments by integrating Earth observation (EO) products, such as GLDAS, GLEAM, and GRACE satellite data, with in situ GWL measurements. Four Machine Learning (ML) models were evaluated for their ability to fill GWL data gaps: Random Forest (RF), XGBoost (XGB), Support Vector Regression (SVR), and Long Short-Term Memory (LSTM). Spatial trend analysis and trend-based extrapolation scenarios of GWL and EO data were performed. New hydrological insights for the region The results demonstrate that RF was the best-performing model, yielding R² values between 0.7604 and 0.9475, NS values between 0.7604 and 0.9460, WI values between 0.7812 and 0.9722, and a minimal RMSE of 0.0210 m. Error analysis indicates high model reliability, with PBIAS values consistently falling within the optimal range of ±10%. XGBoost demonstrated underfitting. The LSTM effectively captured long-term trends despite being susceptible to overfitting. Spatial trend analysis of GWL shows that from 62 wells, 77.42% show a decreasing trend. An exploratory scenario based on historical trends projects severe GWL depletion by 2030 and 2040, reaching −36.75 m in the dry season. This study highlights the potential of integrating downscaled EO data with machine learning algorithms to enhance groundwater assessment in data-scarce regions.