From Pixels to Buildings

This paper describes the land use classification of multispectral digital aerial images, the removal of buildings from the digital surface model and the visualization of the textured digital surface model with reconstructed buildings. The proposed approach applies spectral classification techniques to multispectral digital aerial images with RGB and NIR channels. The results of the land use classification are used both in dense matching and building extraction. Dense matching uses the knowledge of water areas to prevent wrong matches due to non-Lambertian reflections. The obtained digital surface model is used to generate the corresponding RGB ortho images. Further, in building extraction the classification is used as starting point for searching exact building borders and textures for building facades. The buildings in the digital surface model are replaced by refined building models using feature information. Results from a huge test area in city center of Graz are presented and analyzed. This approach produces automatically appealing visualization after a short interactive training phase for classification.