Enhancement of image-to-image co-registration accuracy using spectral matching methods

Two of the important stages in the production of maps from remotely sensed imagery are rectification and classification. Residual geometric errors following rectification produce a “component” of the apparent classification error. Hence, besides being of value in its own right, reduction of the rectification error will enhance the quality of the classification stage, and the classification accuracy statistics will more closely refer to “pure” classification error rather than classification error due to pixel mismatch. In rectification to ground control points (GCPs) there is a natural limit of what can be achieved in rectification of an image, and this is determined by the cost of collecting good GCPs. However, if we are concerned with a sequence of images, and are concerned primarily with estimating change and growth, the rectification is image-to image, and the process becomes one of pixel matching, and the only cost is processing time. We present a spectrally-based pixel-matching algorithm which seems to offer considerable scope for very accurate pixel-to-pixel and hence image-to-image matching. An algorithm is used which maximizes local correlations in each spectral band at each co-registration point: a multivariate anisotropic spatial auto-correlation approach. The results are demonstrated and validated in a case-study of a 1987 Landsat5 TM image, 1997 Landsat5 TM image and a 2000Landsat7 ETM image, all of the same forested region in China.