Critical Evaluation of the SIFT Operator for Remote Sensing Imagery

Abstract

This thesis evaluates the applicability of the SIFT operator for registration of remote sensing imagery. The SIFT operator has been utilized successfully in several fields, but a critical evaluation of its capabilities in the field of remote sensing has not been considered until now. Of special interest is its applicability for radar image registration, where the characteristic scale invariance of the SIFT operator should prove beneficial.

The SIFT operator implementation, presented here, has undergone certain changes to enhance its performance for remote sensing imagery. Improvements have been realized in the preprocessing, keypoint detection and matching phases. A fine matching of the matched points by a suitable similarity metric is also recommended. The additions and modifications have been made keeping radar image registration as the target application.

The SIFT operator has proven its worth for monomodal radar images but might not be useful for multimodal image registration involving radar and optical images. In general, the operator is capable to achieve a pixel level registration consistency for multi-sensor and multi-temporal radar image registration. Compared to mutual information based image registration, the SIFT operator is much faster but might not achieve subpixel accuracy without an additional fine matching of the SIFT matches.