Data driven prediction of uranium in groundwater for environmental radioactivity surveillance.

Uranium occurrence in groundwater is governed by complex hydrogeochemical processes, yet routine monitoring remains analytically demanding. A dataset of 1295 groundwater samples collected from a coastal region of southeastern India during 2016-2025 was used to develop a machine learning framework for predicting uranium concentrations from routinely measured physicochemical parameters. Multi-threshold binary classification was implemented at 2, 15, and 30 µg L⁻1, representing precautionary, guideline, and regulatory levels, respectively. Advanced ensemble learning algorithms were evaluated under pronounced class imbalance conditions. Under pronounced class imbalance, CatBoost achieved the highest performance at 2 µg L⁻1 (F1-score = 80.8%), while LightGBM performed best at 15 µg L⁻1 (F1-score = 66.6%). At 30 µg L⁻1, Isolation Forest achieved 100% recall with a 1.12% false-positive rate, reducing laboratory screening workload by 98.7%. TDS, hardness, chloride, and sulphate were the dominant predictors, with SHAP analysis linking their contributions to mineral dissolution, salinity evolution, and carbonate complexation influencing uranium mobility. Climate change may further modify these processes through altered recharge, water-rock interactions, and groundwater salinity (Barbieri, Marino Domenico et al.,2021). The study uniquely integrates routine physicochemical predictors, multi-threshold classification aligned with evolving WHO guidelines, and one-class anomaly detection for rare-event uranium screening. The framework offers a cost-effective approach for environmental monitoring and early warning, supporting public health protection, regulatory decision-making, and sustainable groundwater.