Characterising urban environments in Sub-Saharan Africa with satellite imagery and unsupervised deep learning

Sub-Saharan Africa is the fastest urbanising region globally, with cities experiencing rapid changes in infrastructure such as roads, buildings, vegetation and land use. Monitoring these developments is crucial for informing measures that make urban development sustainable and benefit city residents. Using high-resolution satellite imagery and unsupervised machine learning, this thesis presents a new framework to characterise urban environments in Accra and three other Sub-Saharan cities. In Accra, the method identified distinct and interpretable urban phenotypes. Clusters captured either single urban features, such as water bodies, or intricate combinations like buildings surrounded by vegetation. With its nuanced analysis of Accra, including sensitivity analyses on scale and cluster number, this approach set a strong foundation for expanding to other cities in Sub-Saharan Africa. In the second part, the method was extended to multiple cities, using a two-step feature learning and clustering approach. It was applied to four Sub-Saharan cities: Accra, Dakar, Dar es Salaam and Kigali. The framework identified shared urban phenotypes across all cities, including densely populated areas and informal settlements both within and outside the city core. It also captured city-specific features such as patterns of urban vegetation and urban agriculture in Dar es Salaam and Kigali. This work revealed common growth patterns across African cities. Specifically, Accra, Dakar, and Dar es Salaam share similar urban features, largely due to their colonial his- tories, even though they differ in climate and historical backgrounds. Kigali differs in its lack of a strong colonial footprint, yet shares other features, such as mixed urban-rural landscapes. This thesis demonstrates that high-resolution satellite images, coupled with unsupervised deep learning provide an interpretable and scalable solution for tracking urban development in near real-time, particularly in low-income countries where traditional data are scarce.