Cross-Regional Transfer Learning for Interpretable Prediction of Groundwater Uranium Contamination

High concentrations of uranium (U) in groundwater pose significant health risks, including nephrotoxicity and cancer, yet predicting U occurrence in data-scarce regions remains challenging. This study develops a cross-regional transfer learning framework based on the TabNet neural network, employing a pretraining-fine-tuning strategy. A TabNet model was pretrained with groundwater U data from California and fine-tuned with limited samples from the Datong Basin, China. The fine-tuned model achieves balanced sensitivity and specificity, enabling basin-wide prediction of groundwater U exceedance probabilities (>30 μg/L) in Datong Basin. In high probability areas (>0.6), more than 99.9% of grid cells show coefficients of variation below 0.5, indicating low epistemic uncertainty. These high risk zones account for 11.4% of the basin. Based on the mean predictions of five submodels, 5% of the population is potentially exposed, with an uncertainty range of 2-12%. The model identifies previously unrecognized U hotspots, providing guidance for monitoring and management. Model interpretation highlights precipitation, slope, and cropping intensity as dominant predictors. Although these predictors are shared between California and Datong, their relative influences differ, reflecting local hydrogeochemical contexts. Overall, these findings demonstrate the feasibility of transferring U-related knowledge across regions and its potential for groundwater quality prediction in under-monitored areas.