Quantifying Watershed Criticality via Deep Learning and Explainable AI for Groundwater Resilience

Özlem Baydaroğlu , Serhan Yeşilköy 

Groundwater is a vital freshwater resource that supports drinking water supply, agricultural sustainability, ecosystem functioning, and drought resilience. As water scarcity and climate change impacts intensify, sustainable groundwater management plays an important role in groundwater management, with groundwater storage (GWS) acting as a key indicator of groundwater resilience. This study introduces an integrated, data-driven framework combining deep learning and explainable artificial intelligence (XAI) to analyze GWS anomalies and quantify watershed criticality across Türkiye. Utilizing Empirical Orthogonal Functions (EOF) for dimensionality reduction of hydro-climatic variables, including GWS, precipitation, and evapotranspiration, a spatiotemporal prediction model based on the Graph WaveNet (GWN) architecture has been developed to capture complex spatial and temporal dependencies. To evaluate watershed-specific vulnerabilities, SHAP (SHapley Additive exPlanations) values have been leveraged to isolate the direct pressure of individual climatic drivers on GWS variations. These predictive and explanatory indicators have been subsequently fused with historical GWS-derived features to establish a multi-criteria watershed criticality score, the robustness of which has been rigorously validated through comprehensive sensitivity analyses. The proposed framework successfully identifies watersheds at imminent risk, offering a scalable projection tool to enhance watershed resilience and support proactive policymaking against climate-induced water scarcity.