An Evolutionary Hybrid Machine Learning Architecture for Groundwater Dynamics Modeling Using Spatial-Temporal Satellite Anomalies

Groundwater management in arid and semi-arid regions faces significant challenges due to the spatial-temporal mismatch of available hydrogeological data and the limited physical interpretability of conventional data-driven models. To address these methodological gaps, this study introduces the foundational architecture of an evolutionary hybrid machine learning framework designed to simulate complex aquifer dynamics. The proposed methodology integrates continuous piezometric records from a highly responsive unconfined aquifer (Nebraska Sandhills, High Plains, USA) with macro-scale Gravity Recovery and Climate Experiment (GRACE) satellite anomalies and high-resolution meteorological variables. To capture macro-hydrological trends and aquifer inertia, data streams are resampled to a synchronized monthly frequency and engineered with temporal lag features.

The computational core features a meta-heuristic stacking ensemble engine. To eliminate human bias in hyperparameter tuning and model aggregation, a Genetic Algorithm (GA) subject to strict mathematical convexity constraints autonomously optimizes the weighting of XGBoost and Long Short-Term Memory (LSTM) base learners. Preliminary validation on chronologically sequestered test data demonstrates the framework's exceptional efficacy. The GA-optimized hybrid ensemble achieved a Coefficient of Determination (R^2) exceeding 0.97 and reduced the Root Mean Square Error (RMSE) to approximately 0.33 ft. This evolutionary approach significantly outperformed standalone algorithms, yielding an error reduction of over 13% compared to the best-performing baseline deep learning model. Furthermore, Explainable AI (XAI) incorporated via SHAP (SHapley Additive exPlanations) values provided quantitative physical validation. The SHAP analysis successfully isolated a distinct 3-month physical recharge lag of precipitation events, proving the network's hydrogeological comprehension. This validated architecture establishes a highly transparent, non-black-box modeling environment, laying the critical groundwork for future strategic forecasting under environmental stress.