Spatiotemporal prediction and attribution of groundwater storage anomaly using enhanced hybrid deep learning modeling...

Large-scale groundwater prediction inherently involves data characterized by strong spatial and temporal heterogeneity as well as high dimensionality. Conventional data-driven approaches often struggle to effectively capture the complex and nonlinear groundwater dynamics, leading to great challenges in terms of accuracy and reliability. To achieve transparent and trustworthy predictions, this study proposes advanced hybrid deep learning modeling methods that integrate groundwater prediction, attribution analysis and uncertainty quantification within a cohesive framework. Two hybrid models with distinct architectures are constructed to effectively capture spatial patterns and temporal dependencies: CNN-Attention-LSTM (CAL) and Transformer-LSTM (TL). Intrinsic contributors are interpreted through SHapley Additive exPlanations. Stein Variational Gradient Descent (SVGD) is incorporated to quantify uncertainty and produce probabilistic predictions by capturing complex posterior distributions. Effectiveness of the methodology is well demonstrated through its application to groundwater storage anomaly (GWSA) prediction across the Yangtze River basin, China. Results show that both CAL and TL models deliver accurate GWSA distributions of spatiotemporal heterogeneity over the basin, with average R² above 0.90. The CAL model outperforms in localized GWSA prediction. Meteorological factors predominantly contribute to GWSA, accounting for 80.66 % in the middle and lower basin. Furthermore, interactions among features and their synergistic impact on GWSA behave sensitive to feature ranges. SVGD significantly enhances predictive reliability, with major observations falling within 95 % confidence intervals. The modeling signifies a promising advancement for accurate GWSA prediction and risk-informed decision-making, while improving interpretability of outcomes. Our method has broad applicability and scalability by flexibly updating feature data in other environmental tasks.