Urban Extent Mapping Using Object-Based Texture Classification and Landsat Data

More than half of the world's population lives in urban areas, and yet urban population is still growing rapidly. It is important to monitor, understand, and model the growth of urban land and population. One of the prerequisites of many urban studies is detailed urban maps. At continental to global scale, Landsat is an ideal source of data for urban extent mapping. However, it is difficult to map urban areas using spectral data only, mainly due to spectral similarity between some urban and non-urban objects and spectral variability within the urban and non-urban class. Here we present an approach for mapping urban extent using object-based texture measures. Preprocessed Global Land Survey (GLS) 2010 Landsat surface reflectance images were segmented using a hierarchical segmentation software package and texture features were extracted at multiple levels of the segmentation hierarchy. We used training data derived from high-resolution imagery to train random forest classifiers for different continents. This method was evaluated in different areas from Europe and North America. Rigorous cross-scene validation gave overall accuracy score of 91.1%. The derived classifier for GLS 2010 images was also tested on GLS 2000 images and generated good results. It is shown that the proposed approach has great potential for global scale urban extent mapping.