Extraction of Slum Areas From VHR Imagery Using GLCM Variance

Many cities in the global South are facing the emergence and growth of highly dynamic slum areas, but often lack detailed information on these developments. Available statistical data are commonly aggregated to large, heterogeneous administrative units that are geographically meaningless for informing effective pro-poor policies. General base information neither allows spatially disaggregated analysis of deprived areas nor monitoring of rapidly changing settlement dynamics, which characterize slums. This paper explores the utility of the gray-level co-occurrence matrix (GLCM) variance to distinguish between slums and formal built-up (formal) areas in very high spatial and spectral resolution satellite imagery such as WorldView2, OrbView, Quickbird, and Resourcesat. Three geographically different cities are selected for this investigation: Mumbai and Ahmedabad, India and Kigali, Rwanda. The exploration of the utility and transferability of the GLCM shows that the variance of the GLCM combined with the normalized difference vegetation index (NDVI) is able to separate slums and formal areas. The overall accuracy achieved is 84% in Kigali, 87% in Mumbai, and 88% in Ahmedabad. Furthermore, combining spectral information with the GLCM variance within a random forest classifier results in a pixel-based classification accuracy of 90%. The final slum map, aggregated to homogenous urban patches (HUPs), shows an accuracy of 88%-95% for slum locations depending on the scale parameter.