An Integrated Machine Learning and GIS Approach for Optimizing Hydrological SCS-CN Parameters and Runoff Prediction in the West Bank

Abstract

Estimating runoff is a practical necessity in semi-arid regions, where a single intense storm can cause flooding in a season that is otherwise short of water. This study looks at the West Bank, where sparse gauging makes rainfall-runoff modelling difficult, and combines the SCS-CN method, GIS and machine learning to get more reliable runoff estimates. We processed rainfall records from 41 stations covering 2020–2024 and used them to build a hydrological database of nine variables, the most important being maximum precipitation (Pmax) and the curve number (CN). Before trusting the data-driven results, we calibrated a HEC-HMS model against observed discharge from the Alexander Basin. Adjusting CN for antecedent moisture condition (AMC) was what allowed the simulated peak discharge to match the observed one. Six machine learning models were then compared: Support Vector Regression (SVR), Random Forest, XGBoost, K-Nearest Neighbours (KNN), Decision Tree and Linear Regression. SVR came out on top with R² = 0.981 and an RMSE of 0.79 mm. Random Forest was second (R² = 0.974), followed by XGBoost (R² = 0.958). Across the models, Pmax turned out to be the strongest driver of surface runoff. These results suggest that machine learning can be a dependable tool for estimating runoff in ungauged basins, and the same approach could be applied to other semi-arid regions for flood risk assessment and water resource management.