Rainwater harvesting potential zone identification using GIS –Machine learning approach and comparison of their performance

Rainwater harvesting (RWH) systems play a vital role in sustainable water management, yet their adoption is often constrained by the difficulty of accurately assessing site suitability and economic feasibility. This study aims to develop a machine learning (ML)–based geospatial framework to map RWH suitability in the Choke Watershed, Ethiopia, using twelve biophysical and hydrological conditioning factors. All spatial datasets were standardized and processed at a 30-m resolution, and the final suitability maps were generated at the same spatial scale to ensure consistent spatial comparability. Three ML algorithms—Random Forest (RF), Gradient Boosting (GB), and Extreme Gradient Boosting (XGBoost)—were trained and validated using 175 suitable and 75 non-suitable inventory points, with a 70/30 training–testing split. Prior to model development, continuous predictors were normalized using min–max scaling, and all datasets were checked for missing values; incomplete DEM-derived pixels were corrected using bilinear interpolation, and unresolved records in thematic layers were removed to avoid bias during training. Model performance was evaluated using confusion-matrix metrics and the area under the receiver operating characteristic curve (AUC-ROC). The GB and XGBoost models showed the highest predictive accuracy (AUC = 0.97), followed by RF (AUC = 0.96). Spatial analysis revealed that highly suitable RWH zones are concentrated in flat terrain with loam and clay-loam soils and higher rainfall availability. This study represents the first application of GB and XGBoost models for RWH suitability mapping in the Choke Watershed and demonstrates the effectiveness of ML-driven geospatial modeling for identifying optimal RWH locations. The findings provide a valuable decision-support tool for planners and policymakers seeking to improve water resource management and climate-resilient agricultural development in the region. • A GIS–Machine Learning framework identifies optimal rainwater harvesting zones. • Twelve biophysical and socio-economic factors integrated for spatial analysis. • Three ML models compared to evaluate RWH prediction accuracy and performance. • The best-performing model produced high-resolution suitability maps for RWH. • Findings support sustainable water management and agricultural resilience planning.