A Hybrid Machine Learning Framework for Spatiotemporal Flood Susceptibility and Risk Forecasting in Dire Dawa, Ethiopia

Abstract

Dire Dawa, Ethiopia, a rapidly urbanizing semi-arid city in the rift valley, experiences recurrent and intensifying flash floods, largely triggered by episodic heavy rainfall along the Dechatu River and its tributaries. This study presents a rigorously validated machine learning framework for high-resolution flood susceptibility mapping and dynamic risk forecasting, designed to strengthen disaster risk reduction and urban resilience planning. Using a random forest ensemble, the model incorporated hydrological (river proximity, flow accumulation), topographic (TWI, elevation, slope), meteorological (48-hour and peak hourly rainfall), anthropogenic (built-up density), and land-cover (NDVI) variables. Performance evaluation demonstrated exceptional results: perfect discrimination (ROC AUC = 1.000), complete flood detection (recall = 1.000), strong spatial overlap (IoU = 0.781 for high-risk zones), and excellent calibration (Brier score = 0.0207). Bootstrap resampling confirmed robust feature stability (>0.85), with 48-hour rainfall, river proximity, and TWI as dominant drivers. High-resolution maps identified concentrated very high-risk zones (4.1–5.2% of the area), including critical hotspots such as Genda Kora and Sabian Lowlands. Time-series simulation of the April 2023 flood accurately captured rapid risk escalation, providing valuable lead-time insights. The framework’s novelty lies in combining near-perfect predictive accuracy, uncertainty quantification, and operational validation in a data-scarce African urban context. Implementation recommendations include immediate deployment, enforcement of riverine buffers, and integration with hydrodynamic models to enhance inundation forecasting and climate-resilient urban planning.