Radar-Textured AI Slashes Slum Mapping Errors in African Cities

Accurate mapping of informal settlements in Sub-Saharan African cities has been a persistent challenge because optical satellite imagery often confuses slums (Local Climate Zone 7) with spectrally similar compact low-rise areas (LCZ 3). Researchers from the University of Pavia and partners introduce a context-aware, reproducible framework that fuses Sentinel-2 spectral bands with Sentinel-1 synthetic aperture radar (SAR) data. They implement a three-tier SAR integration strategy: calibrated backscatter, GLCM texture features, and a physics-guided metric capturing structural disorder and weak radar returns typical of slums. Tested across Nairobi and Eldoret (Kenya) using a stratified hold-out protocol and season-aware ablation, the Optical-SAR model achieves overall accuracy of 0.816 (dry season) and 0.807 (wet season).

The most striking improvement is the reduction of critical LCZ 3–LCZ 7 confusion to 7%, compared to the standard WUDAPT baseline (overall accuracy 0.704). SAR-derived textures provided the dominant performance gain, and seasonal analysis shows that while optical-only separability varies with vegetation phenology, SAR textures stabilize informal settlement mapping across dry and wet seasons. However, the study notes that cross-city transfer remains limited without local adaptation strategies, meaning the method works best when retrained on target cities. This work demonstrates that incorporating radar structural information yields consistent gains for urban morphology mapping in data-scarce environments, offering a practical path for urban planning and SDG monitoring in Sub-Saharan Africa.