Water Scarcity Attribution in the Western U.S. using Interpretable Machine Learning Emulators of Process-Based Hydrologic Models
Date:
Poster: American Geophysical Union (AGU) Fall Meeting, New Orleans, LA
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 Scarcity Attribution in the Western U.S. using Interpretable Machine Learning Emulators of Process-Based Hydrologic Models.” American Geophysical Union (AGU) Fall Meeting. (Poster)
Abstract
Water scarcity in the Western United States results from both meteorological variability and water use practices. Understanding the relative contributions of each is critical for identifying policy interventions to manage water shortage risks. Due to the complexity and interconnectedness of the hydrologic system, macro-scale models are valuable for simulating how physical and socioeconomic pressures increase the risk of water shortages. However, understanding the drivers of water shortage dynamics is challenging due to nonlinear interactions and computational expense. We address this by emulating the UNH Water Balance Model (WBM), using meteorological (precipitation, temperature) and socioeconomic inputs (population density, cropland fraction, irrigation efficiency, and sectoral water-use intensities) to reproduce unmet demand ratios (2010–2019) across Western U.S. Water Management Areas. To move from prediction toward explanation, we apply six complementary Interpretable Machine Learning (IML) methods—R2-based, permutation feature importance, SHAP, weight-based, gradient-based, and LIME metrics—and synthesize results via spatial SHAP distributions, dominant driver maps, feature importance rankings, and consensus‑strength diagnostics. In general, precipitation consistently emerges as the dominant factor in mountainous and northern basins. Among socioeconomic variables, cropland fraction is typically strongest, followed by domestic/industrial water-use intensities and irrigation efficiency; population density ranks lowest across all six methods. Spatially, agricultural extent and water-use intensities dominate in many arid Southwest areas; population density dominates along the Pacific coast. Because individual IML methods perturb inputs differently, their dominant feature maps frequently disagree, resulting in high explanation uncertainty. We construct consensus dominant-feature and consensus-strength maps highlighting where attribution is robust versus uncertain, providing clearer signals than single methods alone. By translating complex model behavior into region- and time-specific driver maps, the approach offers insights for risk-informed planning, targeted demand management, and adaptive water governance.
Citation
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 Scarcity Attribution in the Western U.S. using Interpretable Machine Learning Emulators of Process-Based Hydrologic Models. AGU Fall Meeting Abstracts, 2025, H41V-1500.