An explainable AI approach to deciphering groundwater depth responses to climate variability and human activities in...

Groundwater is critical to water security in the arid and semi-arid Western U.S., yet a comprehensive, large-scale understanding of its response to climate variability and human activities remains elusive. This knowledge gap persists due to sparse observational data, the computational cost of physical process-based models, and complex, nonlinear relationships. To overcome these limitations, the objective of this study is to generate monthly Depth-to-Water maps at a ~ 4 km resolution from 2001 to 2020 for surficial aquifers in the Western U.S. and to explicitly quantify the relative roles of climate versus human factors. This study employs an eXtreme Gradient Boosting machine learning model integrating diverse meteorological, geological, topographical, hydrological, and anthropogenic datasets. Furthermore, an explainable AI technique, SHapley Additive exPlanations, is tailored to interpret the trained model and examine the spatial patterns of climate-driven versus anthropogenic effects on Depth-to-Water changes. The model achieved strong predictive performance (R² = 0.91, RMSE = 36.89 ft) on the independent test dataset (2017-2020). Trend analysis revealed significant groundwater declines in the southern Central Valley, Central and Southern High Plains, and Texas. The tailored explainable AI analysis demonstrated that driver importance exhibits high spatial variability: climate factors primarily control Depth-to-Water dynamics in the Columbia Plateau and Arizona Alluvial aquifers, whereas anthropogenic influences prevail in the High Plains, Central Valley, and Snake River Plain aquifers. Results confirmed that deeper groundwater systems appear more affected by anthropogenic pressures from prolonged extraction and land cover change, while shallower groundwater is more sensitive to climatic variability. Furthermore, climate variability has an indirect effect in accelerating groundwater depletion in heavily agricultural regions by increasing pumping demand. By bridging machine learning predictions with physical attributions, this study provides spatially explicit insights for developing sustainable and targeted water management strategies across the West.