A Stable Land Cover Patches Method for Automatic Registration of Multitemporal Remote Sensing Images

We propose a stable land cover patches method (SLCPM) for the automatic registration of multitemporal remote sensing images. Our method takes advantage of multispectral features as well as stable and widespread land cover patches in remote sensing images. We tested our method using optical satellite images through four different experiments. In the first and second experiments, we tried to register images acquired on different days but by the same sensor. In the third experiment, we tried to register images acquired on different days and by different sensors (which means different spectral resolutions were also considered). In the fourth experiment, we tried to register images with different spatial resolutions acquired on different dates and by different sensors. Three indices were used in our paper for quality evaluation: overall Root Mean Square Error (RMS all), Root Mean Square Error calculated by the leave-one-out method (RMS LOO), and a statistical evaluation of the goodness of control point (CP) distribution across the image (S cat). Results showed SLCPM could generate sufficient, accurate and well distributed CP pairs. We further compared our method with two other popular automatic image registration methods-scale invariant feature transform (SIFT) and a contour-based approach. The contour-based approach could hardly generate any CPs in all the experiments, while SIFT performed very well in the first three experiments both in accuracy and distribution of CPs but was ineffective in the most complex (i.e., last) experiment, due to the lack of correct CPs.