Advances in DSM Generation and Higher Level Information Extraction from High Resolution Optical Stereo Satellite Data

Kurzfassung

The automatic generation of digital surface models (DSM) of urban areas from high and very high resolution (VHR) stereo data from satellites is still a research issue. Image matching algorithms from computer vision have been introduced and adopted to satellite imagery in recent years. These algorithms do not work using local optimisation like area based matching but try to optimize a global cost function. Analysis shows that matching approaches based on epipolar images like semi-global matching (SGM) and new methods using total generalized variation (TGV) yield the best results. Especially satellites like Worldview-2, GeoEye-1 and Pleiades exhibit very high spatial resolution and geometric quality and can therefore be used to generate DSMs with good properties. If several stereo images from one orbit are available, a combination of DSMs generated by different stereo pairs leads to even better results. Using these DSMs, which already show urban area features in high detail, further higher level information can be extracted and new products can be generated. The main focus of this contribution is to present and analyse several derived information products which can be generated using the DSMs generated from satellite stereo data together with panchromatic and multispectral images. In many applications the product of interest is not the DSM but the digital terrain model (DTM), which exhibits not the height of objects like buildings or trees but the underlying terrain. In this paper a selected method and example for generating DTM from DSM is presented. A further issue is to extract real 3D objects like buildings using DSM and the derived DTM data. Different methods have been developed and are presented using typical examples in densely build city areas including validation results. At least larger buildings can be automatically extracted with sufficient accuracy; even the roof shape (gable or flat roof) can be extracted and modelled. Since automatic change detection is generally a very difficult topic in image processing, information on height or 3D properties are of advantage when used in the change detection procedure. Therefore, if stereo data sets from different dates are available, automatic 3D change detection can be performed using the corresponding DSMs. Since the change detection results depend very much on the quality of the DSMs they generally have to be improved using the multispectral information. Two methods have been developed and are shown and compared using examples from dense urban and industrial areas. The presented results show that due to the good data quality and resolution of satellite image data and the corresponding DSMs, it has become feasible to derive higher level and detailed geo-information regarding 3D relevant object classes and change detection.