Automatic Classification of Roof Objects From Aerial Imagery of Informal Settlements in Johannesburg

Improving the planning and provision of public services for those living in informal settlements depends on the availability of accurate demographic information. However, such data frequently do not exist because traditional survey and census methods are rarely successful in these environments. In this paper, the use of automatic feature extraction from aerial imagery is proposed as an alternative to these ground-based methods. We focus on the identification of roof and non-roof objects in an informal settlement called Diepsloot, situated close to Johannesburg in South Africa and home to approximately 200,000 people. Reference data provided by City Johannesburg Metropolitan Municipality authority is used to validate the results of our automated analysis, which achieved an overall accuracy of 80.5 % when compared to manual delineation.