Using satellite images and computer vision to study the effects of spatial apartheid in South Africa
Sefala, Raesetje Bonjo
Removing many of the legacies of Apartheid, a former policy of political and economic discrimination against non-European groups in South Africa, is a primary concern for the country. Aerial images of residential areas show the clear legacy of spatial apartheid, with completely segregated neighbourhoods of townships next to gated wealthy neighbourhoods, a phenomena which has largely remained una ected by the ending of apartheid. This research uses computer vision to analyse 698; 544 satellite images of 9 provinces in South Africa, taking the rst steps toward examining the evolution of spatial apartheid. To achieve this goal, we rst introduce a new dataset consisting of polygons demarcating land use, geographically labelled coordinates of all buildings in South Africa, and high resolution satellite imagery covering the entire country from 2006-2017. Using this dataset, we trained a UNet based semantic segmentation model to detect and classify clusters of buildings for 12 types of classes: Township, Suburb, Industrial area, Commercial land, Informal area, Farm, Collective living Quarters, Village, Smallholdings and Background. We classify these neighbourhoods with an accuracy of 57:45% and a Cohen's Kappa value of 0:4326, giving us the potential to investigate areas a ected by the Group Areas Act which enforced spatial apartheid/segregation.
A dissertation submitted in ful lfilment of the requirements for the degree of Master of Science to the Faculty of Science, University of the Witwatersrand, 2020