Explainable Artificial Intelligence (XAI) Attributions of Water Scarcity in the Western United States are Unstable and Uncertain
Date:
Presentation: American Geophysical Union (AGU) Fall Meeting, San Francisco, CA
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. “Explainable Artificial Intelligence (XAI) Attributions of Water Scarcity in the Western United States are Unstable and Uncertain.” American Geophysical Union (AGU) Fall Meeting. (Presentation)
Abstract
Attributing water scarcity to specific drivers is critical for effective water-resource management but complicated by the spatially heterogeneous and nonlinear interplay of meteorological variability and human water use. We develop a multi-method explainable artificial intelligence (XAI) framework to characterize uncertainty in spatial attributions of water scarcity across the Western United States. We first train a deep-learning emulator using meteorological and human water use inputs paired with simulated unmet water demand generated by the process-based Water Balance Model (WBM). We then apply eight complementary XAI methods to the trained emulator to identify the input variables associated with its predictions.
We find that explanation uncertainty arises from two distinct sources. First, different XAI methods produce substantially divergent attributions when applied to the same trained emulator because they operationalize feature importance differently. Second, even explanations from a single XAI method are highly sensitive to the random seed used to train the emulator, suggesting that the high variance of the emulator can compromise sensitivity analysis. Models with comparable predictive performance can rely on different combinations of precipitation, temperature, agricultural activity, and water usage inputs, reflecting parameter equifinality in the learned representation of the coupled human-water system.
Our findings demonstrate that strong emulation performance does not guarantee stable or scientifically reliable interpretation. XAI attributions can therefore be misleading when training stochasticity and methodological disagreement are ignored. Explicit characterization of explanation uncertainty, together with careful consideration of the bias-variance tradeoff, is essential before model-derived attributions are used to support decisions in coupled natural-human systems.
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. (2026). Explainable Artificial Intelligence (XAI) Attributions of Water Scarcity in the Western United States are Unstable and Uncertain. Abstract GC22D-04 presented at AGU26, San Francisco, CA, 7–11 December 2026.