Machine learning-based identification of shallow groundwater potential zones for sustainable groundwater management in the Guna-Tana landscape, Ethiopia

Study region Guna-Tana landscape of the upper Blue Nile basin, Ethiopia. Study focus Shallow wells are often drilled without proper hydrogeological investigation, resulting in frequent failures. Smallholder farmers and development agents need guiding maps, as trial-and-error well siting consumes significant resources and time. Successful well siting requires delineating shallow groundwater potential zones (SGWPZs) with favorable hydrogeological conditions. This research aimed to identify SGWPZs within the Guna-Tana landscape of Ethiopia. To achieve this, advanced machine learning algorithms (MLAs), Random Forest (RF), Extreme Gradient Boosting (XGB), Classification and Regression Tree (CART), Support Vector Machine (SVM), and K-Nearest Neighbor (KNN) were employed. Groundwater yield data from 1903 hand-dug wells (HDWs) were used as the target variable, and a diverse set of environmental and geospatial features were used as predictors. Each model was trained and validated using 5-fold cross-validation (5-fold CV) within the R-programming environment. New hydrological insights for the region Groundwater potential classified as very high accounted for 7%, 16%, 20%, 7%, and 7% in RF, XGB, CART, SVM, and KNN respectively. Area under the curve (AUC-ROC) scores for RF, XGB, CART, SVM, and KNN were 0.80, 0.78, 0.74, 0.75, and 0.70, respectively. These data imply that RF and XGB are the most reliable models for mapping SGWPZs. The study supports sustainable groundwater management and improves water security in the area by providing policymakers with insightful information.