Automatic building extraction in dense urban areas through GeoEye multispectral imagery

In studies of high-resolution satellite (HRS) imagery, the extraction of man-made features such as roads and buildings has become quite attractive to the photogrammetric and remote-sensing communities. The extraction of 2D images from buildings in a dense urban area is an intricate problem, due to the variety of shapes, sizes, colours, and textures. To overcome the problem, many case studies have been conducted; however, they have frequently contained isolated buildings with low variations of shapes and colours and/or high contrast with respect to adjacent features. As an alternative, this study uses continuous building blocks along with high variation in shape, colour, radiance, size, and height. In addition, some non-building features include either the same or similar materials to that of building rooftops. Thus, there is low contrast between building and non-building features. The core components of the algorithm are: (1) multispectral binary filtering, (2) sub-clustering and single binary filtering, (3) multi-conditional region growing, and (4) post-processing. This approach was applied to a dense urban area in Tehran, Iran, and a semi-urban area in Hongshan district, Wuhan city, central China. A quantitative comparison was carried out between the proposed and three other algorithms for the Wuhan case study. GeoEye multispectral imagery was used in both case studies. The results show that the proposed algorithm correctly extracted the majority of building and non-building features in both case studies. The short running time of this algorithm along with precise manual editing can generate accurate building maps for practical applications.