Abstract

The accelerated growth of the urban population, coupled with a lack of planning, social inequality, a shortage of affordable housing, and high unemployment, has driven the rapid expansion of favelas. The identification and continuous monitoring of these areas are fundamental to supporting the formulation of effective public policies and promoting better living conditions. Traditionally, analyses at more disaggregated geographic scales, such as favelas, are based on census data. However, as this data becomes outdated in relation to its reference date, its ability to adequately represent urban dynamics is reduced. The study proposes an alternative to census data through the application of the U-Net deep learning algorithm for the identification of favelas and urban communities in the city of Rio de Janeiro, using high-resolution RGB satellite images. Orthophotos from 2024, with a spatial resolution of 15 cm, provided by the Pereira Passos Municipal Urban Planning Institute (IPP), were used. The municipality's urban area was subdivided into a regular grid of 512 × 512 meter squares, used to cut the images into smaller units of analysis. The identification of areas of interest was based on the spatial overlap of 2022 census sectors classified as favelas by the Brazilian Institute of Geography and Statistics (IBGE), as well as favela areas mapped in 2019 by IPP. From this overlap, the squares containing areas of interest were selected. The study included four of the municipality's five Planning Areas, from which a stratified sample was constructed using inverse sampling, summing 286 images, corresponding to approximately 75 km². The sample images were resampled using the bilinear method to a resolution of 50 cm. Reference masks were created manually through visual interpretation, considering characteristics associated with precarious housing, such as irregular building patterns, inadequate construction materials, location on slopes, proximity to highways, railways, ditches and open sewers, as well as narrow streets. Of the total images, 80% (229) and their respective masks were used to train two U-Net models, while the remainder was used for evaluation. The models were trained: one without the use of data augmentation and the other with the application of data augmentation techniques, including horizontal mirroring, vertical mirroring and transposition of rows and columns. The model without data augmentation presented an Intersection over Union (IoU) of 0.6842, precision of 0.5189, recall of 0.7751 and F1-Score of 0.6238. The model trained with data augmentation achieved better overall performance, with an IoU of 0.7149, accuracy of 0.6444, recall of 0.7437, and an F1-Score of 0.6592, demonstrating the effectiveness of data augmentation and indicating promising results for the U-NET architecture in the automated identification of favelas in satellite imagery.