A New Approach to Automatic Feature Based Registration of Sar and Spot Images

It has long been known that by combining multi-source or multi-temporal Earth observation images, the ability to detect and recognise features on the ground is greatly increased. In some specific cases models have been successfully developed for automatically integrating images from different sensors, but in general the problem of fully automatic data integration has not yet been solved. This paper introduces a new approach to automatic feature based image registration, based on multiple feature extraction and matching techniques, in order to solve the problem of automatic registration of SAR and SPOT imagery. Application of this model to pairs of small images showed that a substantial number of tie points could be found. Using these tie points with a first order transformation function consistently gave a RMS residual of less than two pixels. The model was also applied to full scene images. A method of processing based on tiling was used, rather than one based on image pyramids, since this was found to be computationally more efficient. Results showed that a large number of tie points could be located across the images. Close inspection of the matched features revealed that there were no blunders and hence the feature extraction and matching algorithm was successful. Most importantly, the approach to feature based registration using multiple feature extraction techniques clearly improved the quality and quantity of the tie points compared to traditional feature based registration techniques which rely on only one feature extraction algorithm.