Automated recognition of urban areas based on land cover composition and configuration

Abstract

Nowadays, urban areas keep growing as a result of rapid economic development, technological advances and population migration. Recognition of an urban area means the identification of the spatial extent of the urban area. It is important to recognize the urban areas of cities, because it provides a basis for classifying urban and rural populations, monitoring and analyzing urban growth, and making governmental decisions and policies. In recent years, increasing availability of remotely sensed data and processing techniques facilitate the development of new approaches to studying urban issues. Remote sensing based approaches have been widely developed for urban land cover / land use classification, urban object extraction and urban landscape analysis. Some efforts have been made to recognize urban areas from remote sensing images, but these methods consider urban area as a thematic class and identify urban areas through a per-pixel classification. These methods do not recognize an urban area as a geographical entity. This research aims to develop an algorithm to recognize urban areas using remote sensing data and techniques. It reviews currently definitions of urban areas to identify common urban characteristics and urban-rural differences from them. Based on the urban-rural differences, relevant information and processes are selected to compose the algorithm. Four urban characteristics are identified through a review of current urban definitions. They are a) urban areas contain large and dense built-up areas; b) urban areas contain heterogeneous elements; c) urban areas are dominant by non-agricultural activities; and d) urban areas are distinguishable from their surrounding rural areas. Eight remote sensing image features are related to the urban characteristics, they are, the four proportions of vegetation, impervious surface, soil and water / shade, and the four textural features including angular second moment, inverse difference moment, contrast and entropy. They correspond to two types of information. Four proportional features correspond to land cover composition, and four textural features correspond to land cover configuration. The experiment results show that the combination of the eight features is valid for characterizing different kinds of areas and effective for distinguishing between urban and rural areas. The multi-resolution image segmentation algorithm is suitable for dividing a city region into homogeneous sub-regions that accord with the physical landscape. In the experiment of the algorithm with Landsat TM data, all the seven spectral bands show a decrease in the average grey-level range along a continuous region splitting process performed for all administrative regions of the study area. The average grey-level ranges in six of the seven bands are further reduced by removing the administrative boundary constraint. An urban area is successfully recognized through an iterative clustering and merging process, performed on the homogeneous regions output from the image segmentation process with the eight proportional and textual features. An experiment shows that the iterative clustering and identification is able to identify an area that can be definitely labelled as urban. Another experiment shows that the iterative merging process is able to identify the urban and rural areas of a city region with the maximum distance between them in the feature space. The resulting urban area is evaluated by a fact consistency checking. By overlapping the resulting urban area with some referenced data, it is verified that all the facts identified about the study area are satisfied by the recognition result.