Data-driven spatial prediction of groundwater levels using machine learning in the gilgel abay watershed, Ethiopia: A comprehensive multi-model evaluation with cross-validation

Abstract

Study region Gilgel Abay Watershed, Upper Blue Nile Basin, Ethiopia (2882 km²) - a volcanic highland catchment where groundwater is an increasingly stressed resource due to population growth and expanding irrigated agriculture under limited monitoring infrastructure. Study focus Five machine learning models - Random Forest (RF), Gradient Boosting (GB), Artificial Neural Network (ANN), Decision Tree (DT), and Support Vector Regression (SVR) - were evaluated for spatially explicit groundwater level (GWL) depth prediction using ten predictors from 122 observation wells (377 observations, 2016–2022), tuned via systematic grid search under 5-fold, 10-fold cross-validation and three train–test split ratios. New hydrological insights for the region RF achieved the best generalization (test R² = 0.82, RMSE = 1.55 m, train–test gap = 0.13) and GB the lowest prediction error (RMSE = 1.30 m, R² = 0.80); both ensemble methods substantially outperformed ANN (R² = 0.75), SVR (R² = 0.64) and, DT (R² = 0.62), confirming ensemble ML as the superior approach for GWL prediction in data-scarce highland watersheds. Permutation-based variable importance revealed lineament density as the dominant predictor - surpassing slope, rainfall, and elevation - establishing tectonic fracture networks, rather than climatic or topographic factors alone, as the primary control on groundwater storage in this volcanic terrain. RF and GB are recommended as operational GWL prediction tools; future work should expand borehole networks and adopt spatially blocked cross-validation for more reliable generalization estimates.