The Shiyang River Basin, an arid inland agricultural watershed in northwestern China, is characterized by intensive irrigation and long-term groundwater extraction that have caused widespread groundwater depletion and increasing pressure on regional water resources. Accurate prediction of groundwater table depth remains challenging in data-limited basins due to complex groundwater dynamics and sparse observations. To address these challenges, this study develops an integrated machine learning framework for monthly groundwater table depth prediction over a 20-year period. The framework combines a Kalman filter for harmonizing heterogeneous datasets, an entropy-based algorithm for identifying dominant influencing factors, and a Time-series Generative Adversarial Network for data augmentation. Spatial characteristics of groundwater table depth are further incorporated to represent the influence of lateral groundwater flow. SHAP-based interpretation reveals that lateral groundwater flow exerts a substantial influence on groundwater dynamics across the basin, highlighting the importance of representing spatial groundwater interactions in data-driven prediction. The proposed framework significantly improves groundwater table depth prediction accuracy compared with conventional machine learning models. This approach provides a robust and interpretable tool for groundwater table depth prediction in arid, data-scarce basins and supports more effective groundwater management in intensively irrigated regions. • We integrated data alignment, data augmentation, and groundwater spatial influence to improve GTD prediction in data-scarce regions. • TimeGAN generates synthetic sequences to augment training data volume and diversity, significantly improving model performance. • Introducing groundwater spatial distribution to characterize lateral flow as physical constraints effectively enhanced the prediction. • SHAP quantified feature contributions and confirmed the key role of lateral flow in prediction for intensive groundwater mining zones.
Yang et al. (Thu,) studied this question.