Water Shortage Attribution in the Western U.S. using eXplainable Artificial Intelligence Emulators of Process-Based Hydrologic Models

Luo, Y., Grogan, D. S., Zuidema, S., Lammers, R. B., Zheng, J., Lisk, M. D., Fisher-Vanden, K., Olmstead, S. M., and Srikrishnan, V. 2025. “Water Shortage Attribution in the Western U.S. using eXplainable Artificial Intelligence Emulators of Process-Based Hydrologic Models.” Cornell Energy & Water Resources Systems Seminar. (Presentation)

Abstract

Water scarcity in the western United States reflects a mismatch between freshwater demand and availability, driven by both meteorological variability (droughts and rising temperatures) and human water use (irrigation and urban and industrial demand). Attributing shortages to specific drivers is difficult because these drivers interact through nonlinearities, lags, and routing. Process-based hydrologic models capture these interactions but are computationally expensive and make it hard to isolate the effect of any single driver, while deep learning emulators are fast but behave as black boxes.

This talk presents a three-stage framework that links the University of New Hampshire Water Balance Model (WBM), a Convolutional Long Short-Term Memory (ConvLSTM) emulator, and an ensemble of six explainable AI (XAI) methods to attribute unmet water demand across 11 western states from 2010 to 2019. The emulator closely reproduces WBM’s monthly unmet demand ratios at a fraction of the computational cost, making large-scale attribution experiments practical. We then apply R²-based leave-one-covariate-out (LOCO), permutation feature importance (PFI), SHapley Additive exPlanations (SHAP), connection weights, input gradients, and Local Interpretable Model-agnostic Explanations (LIME) to the emulator.

Precipitation leads in the Pacific Northwest and Rockies, agriculture in the Central Valley and Plains, temperature effects concentrate in the Southwest, and population density is most important along the Pacific coast. However, different XAI methods often identify different dominant drivers for the same emulator output. Mapping their consensus provides a “confidence map” for attribution, showing where drivers are robustly identified and where attribution remains highly uncertain, particularly in complex, human-dominated basins.

Slides