Interpretable deep-learning DRASTIC for statewide nitrate-vulnerability mapping in Florida's surficial aquifer system

Groundwater is a vital resource that supports drinking water supplies, agriculture, and ecosystems worldwide, but it is increasingly threatened by overexploitation, contamination, and urbanization. Florida is particularly vulnerable due to its shallow Surficial Aquifer System (SAS), karst geology, widespread septic systems, agricultural activities, and climate-related hazards. Groundwater vulnerability maps are important tools for identifying high-risk areas and guiding protection strategies. While deep learning models such as Convolutional Neural Networks (CNNs) and U-Net have improved mapping accuracy by capturing complex spatial patterns, their limited interpretability can reduce stakeholder confidence and practical adoption. This study presents an interpretable groundwater vulnerability mapping framework that combines CNN and U-Net models with the SHapley Additive exPlanations (SHAP) method. The analysis was conducted across Florida using seven DRASTIC parameters: depth to water, net recharge, aquifer media, soil media, topography, vadose zone impact, and hydraulic conductivity. Both models showed strong predictive performance, with AUROC values of 88.52% for CNN and 89.16% for U-Net. SHAP analysis identified depth to water and topography as the most influential factors, with shallow groundwater conditions substantially increasing vulnerability. The results demonstrate the value of explainable AI for transparent groundwater vulnerability assessment and aquifer protection planning.