Impact of Variable Selection on Machine Learning Model Performance for Groundwater Potential Mapping in Western Cameroon

This study evaluates how variable selection affects the performance and robustness of machine-learning models for groundwater potential mapping in the Western Highlands of Cameroon, a hard-rock and hydrogeologically heterogeneous region. An integrated geospatial database was built from 115 boreholes and wells, hydrogeological records, geological and soil maps, ALOS PALSAR radar imagery, Landsat 8 OLI/TIRS imagery, and digital elevation data. Borehole yield was reclassified into groundwater-potential classes using CIEH thresholds and combined with approximately 8,000 pixel-level observations derived from twelve conditioning factors. Eleven supervised classifiers were compared under five feature configurations: all variables, Top 10, Top 7, Top 5, and Top 3 predictors. Model performance was assessed using accuracy, precision, recall, F1-score, MCC, MSE, ROC-AUC, confusion matrices, and class-wise F1-scores. Results show that model accuracy does not increase monotonically with the number of variables. With all features, XGBoost achieved the best accuracy (0.891), followed by CatBoost (0.888) and Random Forest (0.882). With ten variables, Random Forest remained highly accurate (0.877), while CatBoost performed best with seven variables (0.864). Even with five predictors, XGBoost maintained acceptable performance (0.840). These results demonstrate that parsimonious feature subsets can preserve strong predictive power while reducing redundancy, computational cost, and data-collection requirements for operational groundwater mapping.