Investigating the Separability of Slums in Mumbai based on RADARSAT Images using Object-based Image Analysis Methods

Abstract

Global urbanization and the subsequent development of mega cities, like Mumbai in India, drive the emergence of slums. For urban planners, disaster and aid organizations, among others, it is crucial to know the location of slum areas in order to enhance living conditions of the occupants and ensure prompt and organized aid in case of natural disasters. The importance of proper identification of these areas is exacerbated by the high population density and poorly developed infrastructure. In this study, informal (slum) and formal settlements are classified throughout Mumbai using Radarsat-2 imagery. Optical remote sensing imagery has been widely used for urban classification purposes and therefore highly-developed techniques have evolved. On the contrary, the use of Radar remote sensing has not been as prevalent in the past, due to insufficiencies that were better met with the high resolution capabilities of optical remote sensing. However, with the advent of VHR Radar imagery this data source gained renewed interest for urban settlement classification, especially since it is capable of weather- and daytime-independent acquisition, limitations which are present in optical remote sensing imagery. Since techniques for delineating informal settlements in urban environments using Radar imagery have not yet had the maturation time optical data has, they are still in need of further investigation. With the launch of Sentinel-1A in April 2014 and the launch of Sentinal-1B following in 2016, the utility of Radarsat-2 data for informal settlement classification may draw conclusions about the usefulness of the free and widely-available Sentinel-1 data, since specifications are very similar. Various characteristics can be used in order to classify informal settlements. In this thesis several textural features were tested, specifically texture measures derived from Gray Level Co-occurrence Matrices, i.e. Homogeneity, Contrast, Dissimilarity, Entropy, Angular Second Moment, Mean, Standard Deviation and Correlation. Using the texture measures derived for all blocks of buildings of Mumbai, the following three different classification algorithms were investigated: Linear discriminant analysis, Support Vector Machines and Random Forest. The best result was achieved using Random Forest with an Overall Accuracy of 91% and a User’s Accuracy for informal settlements of likewise 91%. Linear Discriminant Analysis showed an Overall Accuracy of 89% but a User’s Accuracy for informal settlements of only 55%, while Support Vector Machines got an Overall Accuracy of 89% and a User’s Accuracy of 64%. All three classifiers underestimated informal settlements to a certain degree.