Last modified: 2016-06-30
Abstract
Due to the accuracy limitation of Attitude and Orbit Control Systems (AOCS) of space-borne remote sensors, the direct geo-location of optical satellite images generally suffers varying degrees of geo-location errors (from several meters to hundreds of meters) on the ground. In practical applications, ground control points (GCPs) are required to perform precise geometric rectification for the optical satellite images. However, the GCPs are not easy to obtain. The traditional GPS-survey approach is time consuming, labor intensive, or even impossible in depopulated areas. An efficient alternative approach to collect GCPs is automatically matching with the reference images, but the high-precision optical reference images are commonly not available without ground survey.
Fortunately, in contrast to space-borne optical images, geo-location of space-borne SAR images is not related to the attitude of space-bore sensor, thus direct geo-positioning without GCPs of space-borne SAR images can reach higher accuracy than space-borne optical image, e.g. absolute geo-location accuracy of TerraSAR-X using science orbit data in 13 worldwide distributed test sites and reaches RMSE values of around 1 m for the Spotlight mode (Bresnahan, 2009). Consequently, it is possible for the space-borne SAR sensors to provide high-precision reference images for the space-borne optical images without any ground labor. Nevertheless, automated geo-referencing between optical images and SAR images is challenging: 1) the optical images and SAR images are commonly quite different in intensity, texture and edges, and 2) the SAR images are seriously contaminated by speckle noises. Both the ordinary area-based (e.g. SSD, NCC) and feature-based (SIFT, SURF) approaches perform poorly for optical image and SAR image matching.
In this paper, we present an efficient and robust approach to automatically matching the optical images with SAR images. 1) To overcome the influence of different sensors and the speckle noises of SAR image, local self-similarity is introduced as the image description instead of SIFT descriptor; 2) both local geometric consistency and global geometric consistency are applied to detect the outliers, and 3) the algorithm is parallelized to process on graphics processing unit (GPU), which can notably shorten the time for automated geo-referencing.