Automatic Building Extraction from High Resolution Stereo Satellite Images

An approach was developed for automatic building extraction from high resolution stereo satellite images. The approach utilizes the spectral properties of the pan-sharpened multispectral bands and the elevation model generated from the stereo panchromatic bands. First, the pan-sharpened multispectral bands are classified using the Maximum Likelihood Classifier (MLC) to separate the buildings from other classes. Next, the Normalized Digital Surface Model (nDSM) is calculated by subtracting Digital Terrain Model (DTM) from Digital Surface Model (DSM). Those areas of nDSM that fall above the user defined threshold are considered to represent the 3D objects. To separate the buildings from the trees, a Normalized Difference Vegetation Index (NDVI) is used. The areas considered to be the buildings on both the classified image and the nDSM are overlaid and the union areas are accepted to represent the candidate building patches. The further processing operations are then carried out using these candidate building patches only. Next, the candidate building patches are applied a Canny edge detector and the detected edges are vectorized using a boundary tracing algorithm, which provides the general boundaries of the buildings. However, the vector boundaries may contain undulations. To remove the undulations therefore, the Douglas Peucker algorithm is used and the topological errors are removed using the vector filters. The approach was implemented in an urban area of Batıkent, Ankara, Turkey using the IKONOS PSM and stereo PAN images. The DTM was generated using the existing 1:1 000-scale digital vector map. The DSM was generated from stereo PAN images and the classification was performed using the PSM bands. The approach was tested using ten urban blocks each containing different types and shapes of buildings. The results show that the proposed approach appears to be quite satisfactory for extracting the buildings with the 80.7 percent detection rate and 72 percent quality rate. The rate of incorrectly labeled building areas was computed to be 0.15, while the rate of missed building areas was found to be 0.24.