Slum mapping in polarimetric SAR data using spatial features

Abstract

Driven by massive urbanization processes, particularly in developing countries, people flock into the cities resulting in an evolution of large slum areas. Mapping and monitoring of slum areas have become an invaluable source for decision-making processes to implement policies related to improve living conditions. Space-borne remotely sensed data has been explored in the past for slum mapping, however, to a large extent supported by optical imagery. In this paper, we explore the capabilities of dual-polarized (HH/VV and VV/VH) X-band Synthetic Aperture Radar (SAR) from TerraSAR-X images for slum extent mapping using the Kennaugh element framework for image preprocessing. In this way, spatial image descriptors based on texture, morphological profiles and polarimetric features have been tested at various window sizes [11 × 11, ... 161 × 161] for mapping slums using the random forest classifier in a series of experiments. For benchmarking the classification results, LDA as parametric linear classifier is used for comparison. Classification performance was evaluated by comparison with a reference map indicating that texture features hold the highest contribution to discriminating slums from other urban structures. Best window size was found using a spatial neighborhood of 81 × 81 pixels resulting in Overall Accuracy of 88.58 and Kappa of 0.7809 for RF classifier. A patch-based analysis of classification results reveals areal dependencies of the classifier in terms of larger slum patches that are mapped with higher precision than smaller patches. Analyses including additional spatial image descriptors based on mathematical profiles reveal no significant contribution to the classification result.