Groundwater nitrate contamination poses a critical public health challenge particularly in arid regions, yet conventional vulnerability indices such as DRASTIC--LU rely on fixed, expert-assigned weights and ratings that limit predictive accuracy across diverse hydrogeological settings and provide no built-in spatial uncertainty quantification. To address these limitations, we developed DD-DRASTIC--LU, a data-driven framework that simultaneously recalibrates both DRASTIC--LU parameter weights and ratings using observed nitrate concentrations across two contrasting hydrogeological domains, while preserving the index's transparent linear structure. The framework derives weights from SHAP importance scores and ratings from partial dependence plots, both extracted from Quantile Regression Forests trained on 324 grid-cell aggregated monitoring wells in Djibouti; uncertainty is quantified through cross-validation variance and interquantile ranges, and integrated into contamination and vulnerability classes through fuzzification, with Shannon entropy quantifying classification confidence at each pixel.

The framework delivers three operational outputs---contamination predictions, specific vulnerability maps, and bivariate risk classifications---and is transferable to other contaminants and hydrogeological settings with model retraining, offering a generalizable protocol for data-driven groundwater vulnerability assessment.