Integrating Socioeconomic Withdrawals Into a Deep Learning Framework for High-Resolution Groundwater Storage Prediction in the Yellow River Basin

Groundwater storage anomaly (GWSA) change provides a vital indicator for assessing groundwater depletion and drought, supporting informed decisions for sustainable water resources management. However, coarse spatial resolution of GRACE data limits detection of local changes, and downscaling often introduces bias by omitting groundwater withdrawals. This study develops a deep learning framework that incorporates socioeconomic groundwater extraction into the groundwater balance to constrain predictions of 1 km resolution GWSA in the Yellow River Basin (YRB), a global representative region under combined stress from climate variability and intensive groundwater use. Results show that the groundwater balance constrained Convolutional Neural Network (CNN) achieves higher accuracy, while the Residual Neural Network (ResNet) better captures spatial variability. Validation against 101 long-term wells (>10 years) confirms 81.2% exhibit correlation coefficients greater than 0.7 with model outputs. Incorporating socioeconomic data into balance constraints significantly improved predictive performance, reducing errors by 28% (CNN) and 68% (ResNet). Groundwater depletion is present across 42.1% of the basin, and within these areas, 92.3% experience high to severe compound stress under drought conditions. These depletion–drought hotspots are predominantly concentrated within 30 km of the main river corridor. While GWSA changes in 60.1% of the basin area are strongly influenced by climate variability and 10.2% by human activities, areas under stronger human influence experience faster depletion (−11.6 ± 0.4 mm/yr) than areas where climate factors dominate (−3.0 ± 0.3 mm/yr). Our study enables high-resolution mapping of groundwater depletion and provides insights for sustainable groundwater management in large dryland basins.