Quantifying Explanation Uncertainty in eXplainable Artificial Intelligence (XAI) Emulators for Water Shortage Attribution in the Western U.S.

Luo, Y., Grogan, D. S., Zuidema, S., Lammers, R. B., Zheng, J., Lisk, M. D., Fisher-Vanden, K., Olmstead, S. M., and Srikrishnan, V. 2026. “Quantifying Explanation Uncertainty in eXplainable Artificial Intelligence (XAI) Emulators for Water Shortage Attribution in the Western U.S.” Cornell Biological & Environmental Engineering 17th Annual Research Symposium. (Poster)

Abstract

Water scarcity in the Western U.S. arises from coupled interactions among hydroclimatic variability, human water use, infrastructure operations, and institutional constraints, making attribution of shortages to specific drivers challenging. Process-based hydrologic models can represent these dynamics, but their computational cost limits large-scale sensitivity analysis and their complexity complicates interpretation. This study presents an integrated framework that combines a deep learning emulator of the UNH Water Balance Model (WBM) with an ensemble of six eXplainable Artificial Intelligence (XAI) methods to characterize the dominant drivers of water shortages and the uncertainty in those explanations. We emulate monthly unmet water demand ratios across Western U.S. Water Management Areas during 2010–2019 using a Convolutional Long Short-Term Memory (ConvLSTM) network trained on meteorological and socioeconomic inputs. The emulator reproduces WBM outputs with high test-set skill, enabling rapid regional attribution analyses that would be computationally expensive with the original process-based model. 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 estimate dominant drivers of unmet demand. We find that different XAI methods can produce substantially different attributions for the same emulator output, with the largest disagreement in mixed urban-agricultural basins. While precipitation is consistently identified as the dominant driver in the Pacific Northwest and Rockies, the relative importance of temperature and socioeconomic variables is highly method sensitive. To address this, we introduce a consensus attribution framework that maps both the plurality dominant driver and cross-method agreement strength, identifying regions of robust attribution and regions of high explanation uncertainty. This work demonstrates how quantifying explanation uncertainty transforms black-box emulators into transparent, reliable tools for scientific attribution in complex human-natural systems.

Poster