Explainable Artificial Intelligence (XAI) Attributions of Water Scarcity in the Western United States are Unstable and Uncertain

Abstract

Explainable Artificial Intelligence (XAI) methods are increasingly used to understand the predictions made by machine learning (ML) models in fields such as hydrology. However, there is a lack of clear guidance about which methods are most appropriate for a given analysis. These complexities result in uncertainties about explanations. We illustrate these uncertainties using a synthetic ground-truth benchmark and a perfect model experiment for attributing water scarcity across the Western United States. Identifying drivers of water scarcity is critical for water-resource management but complicated by interactions between meteorological variability and human-system uses. We train an ML emulator on simulated unmet demand from a process-based hydrological model, then use eight XAI methods to interpret its predictions. We identify two distinct sources of uncertainty: explanation uncertainty from different XAI methods and numerical instability from explanation sensitivity to training seed. These findings show that XAI methods should be selected intentionally and explanations tested for uncertainty before being relied upon.