IMAGE MATCHING ERROR DETECTION WITH FOCUS ON MATCHING OF SAR AND OPTICAL IMAGES

摘要

In this paper, we present matching of multitemporal optical and SAR (Synthetic Aperture Radar) images of various spatial resolutions and naturally different bands. The aims of such matching could be various, like image co-registration, needed for various applications, where a combination of SAR and optical images is needed, as. e.g. in disaster and hazard monitoring and mapping. Here, the application idea is to transfer GCPs (Ground Control Points) from SAR to optical images. SAR images, at least with the newest satellite platforms, can be georeferenced without the use of GCPs down to an accuracy of few dm, while optical sensors, especially high-resolution ones, often need GCPs to achieve the highest possible accuracy. The SAR images are assumed to be orthorectified and a related DSM (Digital Surface Model) exists. Thus, through matching, we can automatically determine planimetric coordinates of GCPs for optical images, which are needed for an accurate optical image georeferencing, while the height is interpolated from the given DSM. However, matching SAR and optical images is so difficult that the great majority of matching results are wrong (only 1 out of about 1,000 match points is correct). In these investigations, we used LSM (Least Squares Matching), as it is very accurate, but most importantly it can provide different statistics, which can be used for matching error detection and elimination. Thus, we developed novel methods for an LSM matching quality characterization and error detection. The same methods, with some minor modifications, can be used for quality characterization for various other matching methods, like cross-correlation, or even feature-based matching. For matching of SAR and optical images, very variable test datasets were used and the results are based on visual inspection of the match-points, very promising, with poorer performance for low spatial resolution images of about 25 m GSD (Ground Sampling Distance). The same matching error detection method has been used for a much more important application, the automatic DSM generation via image matching. Preliminary results show that performing such tests after matching lead to a much more accurate DSM (error standard deviation decreased from 9.3 m to 1.3m from no error detection to the method proposed below). They also provide an accuracy measure for each DSM point, which could be used for various purposes, among them DSM fusion.