A transferable and interpretable approach to slum mapping using building morphometrics and optical imagery

Slums are defined at the household level by deficiencies in housing and basic services, and their identification is central to understanding and addressing urban deprivation. Previous studies relying on very-high resolution imagery and deep learning method often involve costly data acquisition, intensive computational requirements, and limited transparency in model interpretation. To address these challenges, we propose a building-level slum mapping framework that directly classifies individual buildings using a Random Forest model. The framework leverages explicitly semantic morphometrics from open building footprint data, complemented by spectral, proximity-based, and topographic features, all based on publicly available sources. The model was trained and validated on labeled data from over 250,000 buildings across four major cities in Kenya. Under K-fold cross-validation, the full-feature model achieved strong performance (F1 score = 0.987), compared to 0.836 when using morphological features alone. Spatial cross-validation further demonstrated that combining morphological and spectral features yielded the highest average F1 score (0.742), indicating stable generalization to unseen cities. These findings highlight the value of building-level morphometrics for cost-effective and transferable slum mapping. To support broader applications and reduce the risk of stigmatizing individual households, predicted slum buildings are aggregated into 100-meter grid cells, providing a scalable basis for urban vulnerability assessment and sustainable urban planning.