Mapping urban slum settlements using very high-resolution imagery and land boundary data

Abstract

Accurate mapping of slums is crucial for urban planning and management. This article proposes a machine learning, hierarchical object-based method to map slum settlements using very high-resolution (VHR) imagery and land boundary data to support slum upgrading. The proposed method is tested in Kingston Metropolitan Area, Jamaica. First, the VHR imagery is classified into major land cover classes (i.e., the initial land cover map). Second, the VHR imagery and land boundary layer are used to obtain homogenous neighborhoods (HNs). Third, the initial land cover map is used to derive multiple context, spectral, and texture image features according to the local physical characteristics of slum settlements. Fourth, a machine-learning classifier, classification and regression trees, is used to classify HNs into slum and nonslum settlements using only the effective image features. Finally, reference data collected manually are used to assess the accuracy of the classification. In the training site, an overall accuracy of 0.935 is achieved. The effective image indicators for slum mapping include the building layout, building density, building roof characteristics, and distance from buildings to gullies. The classifier and those features selected from the training site are further used to map slums in two validating sites to assess the transferability of our approach. Overall accuracy of the two validating sites reached 0.928 and 0.929, respectively, suggesting that the features and classification model obtained from one site has the potential to be transferred to other areas in Jamaica and possibly other developing Caribbean countries with similar situation and data availability.