Capturing and characterizing the urban fabric of sub-Saharan Africa with very high-resolution satellite imagery and unsupervised machine learning

Keywords: Satellite imagery, remote sensing, deep learning, computer vision, unsupervised machine learning, built environment, urban growth, urban development
Abstract Type: Paper Abstract

Abstract

The spatial signatures of cities can tell us different stories of history, and society, which are all encoded in the composition of the built and natural environment. The urban fabric does not only reflect the past, but also tells us about the present, such as how people are connected to hubs of trade and transport, what environments they are exposed to, and how they live, among other things. Meanwhile, cities in sub-Saharan Africa are growing rapidly, and data on these cities are still scarce and not frequently updated. New methods are needed to fill this gap. 
We present a framework to cluster high-resolution satellite images of multiple cities in sub-Saharan Africa with a deep-learning feature extraction and clustering algorithm. We compare the results of clustering the cities individually and jointly and report city-specific clusters as well as shared clusters between cities. We further contextualize the resultant clusters with demographic and environmental data that were not used for clustering.
We show that image-based clusters captured distinct features of the urban built environment, vegetation, and population. Clusters that consisted of single defining characteristics (e.g., densely populated areas or dense vegetation) were picked up in both approaches, whereas those based on a combination of defining characteristics (e.g., buildings surrounded by vegetation) changed depending on the clustering approach. The results demonstrate that satellite images can be used in combination with unsupervised machine learning to capture and characterize the urban environment. It is a cost-effective and scalable approach for tracking urban development in near-time.