Integrating SVR and STAR-BME for groundwater level forecasting and spatiotemporal mapping in the Choushui River alluvial fan, Taiwan

Abstract Groundwater is an essential water resource in the Choushui River alluvial fan, Taiwan, where intensive groundwater abstraction has caused severe land subsidence and groundwater management challenges. Accurate groundwater level forecasting and regional groundwater assessment are therefore important for sustainable groundwater management. This study integrates support vector regression (SVR) and Space-Time Analytics and Rendering tool–Bayesian Maximum Entropy (STAR-BME) to investigate groundwater level dynamics in the Choushui River alluvial fan. Monthly groundwater level data from 11 monitoring stations, rainfall data, and groundwater pumping data were used to develop groundwater level forecasting models. Four SVR models with different combinations of groundwater level, rainfall, pumping, and seasonal variables were developed and independently validated. Results showed that the model incorporating groundwater level, rainfall, pumping, and seasonal information achieved the best forecasting performance. Forecasting performance gradually decreased as lead time increased from one to three months; however, satisfactory predictive skill was still maintained for short-term groundwater level forecasting. The results also revealed substantial spatial variability in groundwater predictability among monitoring stations. In addition, STAR-BME was applied to reconstruct regional groundwater level distributions and characterize the spatiotemporal variability of groundwater conditions across the alluvial fan. The reconstructed groundwater level maps successfully captured regional groundwater patterns and seasonal variations. The proposed framework provides a practical approach for groundwater level forecasting and regional groundwater assessment and can support groundwater resource management in heavily exploited alluvial aquifer systems.