Author(s): Nebiyou Kassahun; Julian Koch; Xin He; Yicheng Wang
Linked Author(s): Xin He, Yicheng Wang
Keywords: No Keywords
Abstract: Groundwater depletion due to overdraft has been a pressing challenge in the Tangshan Plain, a geologically complex coastal region located in northern China. Existing large-scale assessments were unable to adequately distinguish exploitable freshwater at the local scale or capture spatiotemporal variability for management purposes. This study addresses these issues by developing a machine learning-based spatiotemporal modeling framework in the multi-aquifer system setting of Tangshan. A random forest regressor was trained based on the groundwater level time series records of multiple wells, each treated as an indivisible analytical unit, and both static and dynamic features were used as covariates. Cross-validation explained approximately 0.62 of the variances. SHAP analysis revealed that elevation, sand content, and hydrogeology were the leading spatial drivers, while lagged precipitation and irrigation were the dominant temporal controls. The observed-predicted time series comparison at the 13 test wells yielded a correlation coefficient of r = 0.934 and an RMSE = 8.989, with all predictions falling within the 90% prediction confidence interval. Spatial predictions of the groundwater head at the wet and dry seasonal extremes showed a maximum seasonal head change of 8 m, and the mid-term groundwater prediction showed a mild recovery trend with some localized declines. The results demonstrate that the framework can provide continuous, high-resolution tracking of groundwater head in complex coastal aquifers. Such an approach also enhances the local understanding of exploitable freshwater dynamics, discovers potential areas under abstraction stress, and offers a transferable tool for groundwater monitoring and adaptive management.
Year: 2026