TriFusion-Karst: A spatially robust and hydrogeologically validated stacked-ensemble model for groundwater potential assessment

Abstract Groundwater potential mapping in semi-arid karst regions is often compromised by spatially optimistic validation, limited assessment of predictive variability, and insufficient independent hydrogeological verification. This study introduces TriFusion-Karst, a stacked-ensemble model using RF and LightGBM as base learners and XGBoost as the meta-learner, embedded within a spatially robust and hydrogeologically validated workflow. Thirteen geo-environmental predictors were extracted at 7,313 groundwater monitoring wells, and GTD observations collected during 2018–2024 were pooled and used exclusively as the regression target. TriFusion-Karst and six alternative standalone and ensemble configurations were evaluated using five-fold spatially blocked cross-validation with leakage-controlled out-of-fold meta-feature generation, followed by 30-m spatial prediction. TriFusion-Karst achieved the best performance (RMSE = 1.47 m, MAE = 0.99 m, R² = 0.930, NSE = 0.926, and r = 0.964). Fold-wise spatial prediction variability was quantified using the pixel-wise standard deviation of five fold-specific predictions as an indicator of model stability rather than total predictive uncertainty. Independent validation using 250 mapped karst springs excluded from model development showed that 91.3% occurred within the High and Very High groundwater-potential classes, which covered 42.6% of the study area (SRI = 2.14; p < 0.001; 1,000 permutations). Relief energy, evapotranspiration, fracture density, drainage density, and lithology were the most influential predictors. The resulting maps provide a reproducible basis for regional screening and preliminary groundwater prioritisation but represent relative groundwater favourability rather than aquifer productivity or sustainable yield.