Automatic urban feature extraction using rule-based object-oriented classification: a case study of parts of Pune city, Maharashtra, India

Abstract

Urban areas are gaining attention globally with the implementation of the United Nations sustainable development agenda 2030 where more emphasis is given on making cities inclusive, resilient, safe, and sustainable. Hence, it is crucial to have precise data of urban built-up areas such as the shape, size, and spatial context. It is a challenging task to extract urban built-up features due to continuous modifications in land as well as heterogeneity in spatial and spectral extent of the urban surfaces. The present research attempts to extract urban built up structures using rule-based object-oriented classification. SEaTH, a tool used for feature analysis in eCognition software was applied to select the discrete features and optimum thresholds that allow more and more separability during classification. With respect to diversity in urban areas, two urban patches of Pune city were selected where one patch is the core part of the city with a congested network of roads and buildings and another patch is located in the outskirts comprises of modern multi-story buildings and relatively broad roads. Multiresolution segmentation with scale parameter of 5 with a shape 0.1 and compactness of 0.5 was finally accepted after a lot of trial iterations for both the areas. Using the SEaTH tool, some of the best object features such as shape properties, spectral bands, and indices (NDVI) were selected for the assessment of the separability and threshold. A rule-based classification was performed to acquire land use/land cover with an overall accuracy of 92% for the city core and 91% for the suburb. The k-hat value obtained was 0.81 and 0.88 for the city core and suburb area, respectively. With incorporating shape parameters in image classification, the SEaTH method applied hierarchically the shape features such as density, compactness, and shape index as the best features to separate the buildings and roads. The NDVI spectral index demonstrated in this study proved beneficial to classify vegetation features from other land use types. As a result of the present study, it has been concluded that rule-based object-oriented classification can help improve the classification of dynamic urban areas and update land use maps effectively.