Application of semi-supervised models for groundwater level simulation in arid regions with small sample sizes

Abstract

Accurate groundwater level monitoring is critical for sustainable aquifer management, yet data scarcity in arid regions severely limits predictive capabilities. Conventional models rely on extensive labeled datasets to link meteorological-hydrological variables with groundwater dynamics, a method that is often ineffective in arid basins with sparse monitoring networks. This study introduces a semi-supervised learning model that leverages limited groundwater observations (n = 3150) and unlabeled environmental covariates to predict spatial groundwater levels in the Tarim River Basin (TRB). Compared to traditional groundwater prediction models, the semi-supervised architecture achieved superior performance (R² = 0.87; RMSE = 0.75 ± 0.03) by synergizing adversarial pseudo-label generation and variational feature extraction. SHAP analysis was employed to interpret the model, identifying distance to the river channel (dist_R), hydraulic conductivity (K), Groundwater Storage Anomalies (GWSA), and the Normalized Difference Vegetation Index as key predictors of groundwater level changes. Importantly, downscaled GWSA provided essential insights into the spatiotemporal variations of the groundwater system. The study revealed that the most severe groundwater depletion occurred in the northern TRB, where only 4 % of monitoring wells showed rising water levels, while over 60 % recorded declines exceeding 10m. In the western (70 %) and central (60 %) TRB, groundwater levels also declined to varying extents, though generally within 10m. In contrast, the southern TRB exhibited a rising trend, with approximately 54 % of wells showing an increase in water levels. Despite localized recovery, groundwater resources across the watershed continue to face pressures from over-extraction. Our semi-supervised learning framework demonstrates that physics-informed pseudo-labeling can overcome data scarcity barriers, providing a transferable toolkit for data-sparse arid regions and enabling actionable insights for balancing water allocation and ecological preservation under climate uncertainty.