SlumMapVisionBR: A Dataset for Mapping Slums with Medium-Resolution Satellite Imagery

In this paper, we introduce SlumMapVisionBR, a new dataset composed of freely available medium-resolution images of over one hundred municipalities in Brazil and ground-truth data with slum areas for each location. To allow comparison with future studies, we benchmark this dataset utilising two machine learning methods: multispectral data paired with a random forests classifier and deep learning convolution neural networks. We report our results using common metrics used in slum mapping: accuracy and mean intersection over union. Our experiments show that deep learning performed, on average, better than multispectral data classification when both were evaluated using the same experimental setting. The average accuracy and mean intersection over union were 0.89 and 0.47 for the deep learning method and 0.77 and 0.42 for multispectral data classification. Our dataset and the code to reproduce results have been included in this paper submission and will be made available online upon acceptance of the paper.