Mapping Rio de Janeiro's favelas: general-purpose vs. satellite-specific neural networks

TL;DR

This study addresses the urban informality mapping challenge by comparing task-specific satellite-pretrained networks (Remote Sensing Foundation Models) with large-scale generic pretrained networks for detecting Rio de Janeiro favelas. Using a 150 m tile grid, labeled data, and a random-forest classifier, the authors balance classes and evaluate via 5-fold cross-validation with Precision, Recall, and F1 metrics. The satellite-focused RSFM, CROMA, consistently outperforms generic nets (≈0.81 vs ≈0.72 in F1), demonstrating the advantage of domain-specific pretraining for intra-urban classification even when RGB inputs are used. The findings suggest broader applicability and lower development costs for RSFM-based urban mapping across different informal settlements and cities, leveraging freely available satellite data for dynamic monitoring.

Abstract

While deep learning methods for detecting informal settlements have already been developed, they have not yet fully utilized the potential offered by recent pretrained neural networks. We compare two types of pretrained neural networks for detecting the favelas of Rio de Janeiro: 1. Generic networks pretrained on large diverse datasets of unspecific images, 2. A specialized network pretrained on satellite imagery. While the latter is more specific to the target task, the former has been pretrained on significantly more images. Hence, this research investigates whether task specificity or data volume yields superior performance in urban informal settlement detection.

Figures (1)

• Figure 1: Favela and non-favela tiles. Non-retained tiles are not shaded.