SAR Image Feature Analysis for Slum Detection in Megacities

Abstract

Urban areas have undergone major changes in the last decades. Especially in developing countries, megacities with more than 10 million inhabitants have developed. The number of such huge urban agglomerations is predicted to rise to 41 until 2030. The increasing influx of urban dwellers inevitably leads to the formation of informal settlements with poor living conditions. Monitoring, analysis and mapping of slums are necessary to provide information about such settlements and thus tackle these increasing challenges. Therefore, SAR remote sensing images are used in this study for an extensive mapping of slums conducted in Mumbai, India. In a broad experimental setup textural and morphological features are analyzed for the slum discrimination in an urban landscape. Two state-of-the-art supervised classification algorithms are compared in the experiments. By utilizing an area-wide reference data set, detailed accuracy assessment techniques are applied to determine the classification quality. The results indicate that it is possible to classify the urban landscape by using textural image features with an Overall Accuracy of 88.58% utilizing a Random Forest classifier. However, lower class specific accuracies of the slum areas show that slum mapping remains challenging. A patch based accuracy assessment proves that it is most difficult to detect small slum areas in the urban landscape. Furthermore, an experimental feature reduction experiment indicates that more image features are needed to detect slums than to discriminate urban and non-urban landscapes.