Flood forecasting and early warning In Nyamwamba river catchment using machine learning

Abstract

The Nyamwamba River Catchment in Kasese District experiences frequent, destructive flash floods driven by steep topography, heavy rainfall, and climate variability, exposing critical vulnerabilities in regional disaster management due to sparse gauging infrastructure and reactive response mechanisms. This study develops an integrated machine learning-based flood forecasting and early warning framework to quantify flood risk in the Nyamwamba catchment and to provide timely, location-specific warnings for vulnerable downstream communities. Predictive modeling using Extreme Gradient Boosting (XGBoost)—incorporating binary classification for flood occurrence and regression for river discharge prediction—was forced with multi-source meteorological and remote sensing data (ERA5-Land baseline 2000–2025 and Open-Meteo APIs), utilizing engineered hydrological features such as antecedent precipitation indices, rolling statistics, and flash-flood threats. Spatial flood routing and overland flow simulations were further conducted within an ArcGIS environment to evaluate catchment response times and critical lead intervals under high-intensity precipitation scenarios. Outputs included optimized XGBoost classification and regression models, a real-time operational forecasting pipeline, GIS-based spatial flood simulations demonstrating sub-hourly catchment response times, and an automated multi-channel early warning platform capable of dispatching instant SMS and email alerts to at-risk stakeholders.