On the power of CNNs to detect slums in Brazil

Abstract

The rapid expansion of slums poses a critical challenge for urban planning in Low- and Middle-Income Countries (LMICs), where traditional data collection methods like censuses are often outdated and insufficient. This study examines the transferability and generalization capabilities of deep learning models, specifically Convolutional Neural Networks (CNNs), for automated slum detection across six Brazilian cities with varying urban morphologies: São Paulo, Rio de Janeiro, Belo Horizonte, Brasília, Salvador, and Porto Alegre. Utilizing Very High Resolution (VHR) and High Resolution (HR) satellite imagery, we trained and evaluated models based on the EfficientNetV2L architecture. Our experimental results show that CNN models trained on data from a single city achieved high accuracy within that city (F1 scores exceeding 0.90 with VHR imagery), but their performance significantly decreased when applied to other cities (F1 scores dropping below 0.80), highlighting the impact of regional variations in urban morphology. Conversely, a generalized model trained on combined data from all six cities maintained robust performance across all cities, achieving F1 scores above 0.80 with VHR imagery. These findings indicate that while CNNs are effective for automated slum mapping, regional diversity necessitates training on diverse datasets to ensure generalization. We provide a comprehensive methodology over an openly shared dataset, and code to facilitate future research and applications in urban geoscience. The aim is to enhance the scalability and generalization of remote sensing and deep learning methods for slum identification across diverse urban environments.