Uncertainty Analysis for Image Interpretations of Urban Slums

Image interpretations are used to identify slums in object-oriented image analysis (OOA). Such interpretations, however, contain uncertainties which may negatively impact the accuracy of classification. In this paper, we study the spatial uncertainties related to the delineations of slums as observed from very high resolution (VHR) images in the contexts of Ahmedabad (India), Nairobi (Kenya) and Cape Town (South Africa). Nineteen image interpretations and supplementary data were acquired for each context by means of semi-structured questionnaires. Slum areas agreed upon by different experts were determined. Uncertainty was modelled using random sets, and boundary variation was quantified using the bootstrapping method. Results show a highly significant difference between slum identification and delineation for the three contexts, whereas the level of experience in slum-related studies of experts is not significant. Factors of the built environment used by experts to distinguish slums from non-slum areas or leading to deviations in slum identification are discussed. We conclude that uncertainties in slum delineations from VHR images can be quantified successfully using modern spatial statistical methods.