Analysis of Baseline and Novel Boosting Models for Flood-Prone Prediction and Explainability: Case from the Upper Drâa Basin (Morocco)

Abstract

Flooding poses significant challenges in semi-arid regions, where irregular rainfall patterns increase environmental vulnerability. This study explicitly aims to improve flood susceptibility mapping by integrating advanced machine learning (ML) algorithms with geographic information systems (GIS) and remote-sensing data. Using data from the Upper Drâa Basin in southern Morocco, we applied boosting algorithms, including XGBoost, CatBoost, LightGBM, and Hist Gradient Boosting, to enhance the accuracy of flood risk assessment. Quantitative model evaluation shows that Hist Gradient Boosting achieved the best performance, with the lowest mean squared error (MSE = 0.06897) and root mean squared error (RMSE = 0.2626). It also attained the highest F1 score (0.8), overall accuracy (93.1%), and area under the curve (AUC = 0.833), indicating its superior predictive capability. These findings highlight the strong potential of novel boosting ensemble learning methods in flood susceptibility prediction and contribute valuable, data-driven insights for policymakers and urban planners to support effective flood mitigation strategies in southern Morocco.