Large area mapping of urban deprivation from Sentinel-2 and Google Open Buildings using deep learning

This study explores the potential of a synergistic approach combining Sentinel-2 data and Google Open Buildings (GOB) for mapping urban deprivation over large areas at a 100m spatial resolution. Urban deprivation, including slums, is a crucial aspect of Sustainable Development Goal 11 (SDG): Make cities and human settlements inclusive, safe, resilient and sustainable. Using a Convolutional Neural Network (CNN) and a VGG19 architecture, we experimented with pre-training, self-training, and post-processing using OpenStreetMap to improve classification accuracy. Our results show that combining outputs from both Sentinel-2 and GOB models improves the overall model performance with an F1 score of 81%. The incorporation of post-processing is useful for the final map creation, particularly in correcting for mis-classifications in areas with obvious morphological similarities to deprived areas, such as markets. This approach, which reduces known errors in the model, holds promise for advancing the precision and reliability of urban deprivation mapping over extensive geographical areas.