Mapping Slums from Satellite Imagery using Machine Learning

The United Nations Habitat estimates that over a billion people live in slums; however, determining the location of these communities and keeping up-to-date records of them is still a challenge. Currently, most of the information about slums comes from census data, which is only available at aggregated levels and updated at lengthy intervals, and thus not suitable to capture the dynamism of slums. Consequently, researchers have developed alternative technical approaches to mapping these areas, such as processing satellite imagery to detect and estimate the population living in slums across the globe. Yet, most approaches to date rely on high-resolution satellite imagery, which is expensive to acquire and process, and, as a result, studies tend to be limited to small areas. Moreover, most models developed require large amounts of labelled data, which is presently not available. Additionally, data and models are not publicly available, making it challenging for the scientific community to integrate new findings and convert ideas into practice. This thesis aims to contribute to the development of a technical solution to develop a global slum inventory that can be used to understand and promote sustainable development. In Chapter 3, we show that freely available Sentinel-2 imagery can be used to accurately detect slums in locations in Colombia, Sudan and Nigeria. In Chapter 4, we propose a novel pipeline for identifying slums using feature extraction with deep learning convolutional neural networks and clustering, an unsupervised learning method. In Chapter 5, we introduce SlumMapVisionBR, a novel dataset for mapping slums with Landsat imagery with a thirty-metre resolution that consists of data from over a hundred municipalities in Brazil that contain slums. In Chapter 6, we show that gridded population datasets can be a valuable tool to estimate the population living in slums at the city-scale level. The experiments conducted in this thesis show that freely available medium-resolution satellite imagery can be used to accurately detect slums, hence allowing the study of larger areas. Also, the possibility of employing unsupervised learning could allow the slum inventory to cover areas where there is no labelled data to train models to detect slums, which is the case in most locations with slums around the world. Additionally, being able to inexpensively estimate the population living in slum areas at a city-level scale with gridded population data is valuable for many stakeholders involved in promoting sustainable development. Our dataset, SlumMapVisionBR, is currently the largest publicly available dataset for training machine-learning models to detect slums in cities and constitutes an important step towards making slum mapping research more openly accessible. Also, the code for our is available online. The research presented in this thesis can lead to a cost-effective and scalable end-to-end process to detect and quantify the population living in slums, important requirements for the development of a global slum inventory that can be used to keep track of Goal 11.1 of the Sustainable Development Goals.