Validation of Machine Learning Algorithms for Mapping Flood Inundation and Slum Expansion in East African Cities

Abstract In Sub-Saharan Africa, approximately 50% of the urban population resides in slums. Yet studies examining the spatiotemporal development and flood vulnerability of slums remain scarce in East African cities. This study aims to (i) analyse spatial-temporal dynamics in the Kibera (Kenya) and Katanga (Uganda) slums, comparing with Bannyahe (Rwanda), whose residents were successfully relocated to Busanza model village, and (ii) to quantitatively assess flood exposure across various land use/land cover (LULC) categories and exposed population. We applied the Random Forest (RF) Algorithm classification to Landsat 7,8 and 9 (2012–2024) for LULC mapping, alongside Synthetic Aperture Radar (SAR)-based flood detection using Sentinel-1. The model was trained with a 70% training and 30% testing split and evaluated via confusion matrices and Kappa, achieving 92.75% accuracy for RF and 89% for SAR. The results show (1) that built-up areas increased significantly in Kibera (86.67%) and Katanga (77.52%) slums. In Bannyahe experienced a decline of 32.22% after 2021, which is attributed to the successful relocation to Busanza model village. (2) An overlay analysis of flood exposure across LULC categories during all flood events revealed total flooded areas of 112.50 hectares in Kibera slum and 7.20 hectares in Katanga slum. Population estimates indicate that more than 45,000 residents were exposed in Kibera. Importantly, our algorithms produce flood maps over affected areas and estimate exposed populations within minutes, making them easily applicable to other regions. This study recommends a similar resettlement approach adopted in Bannyahe.