Multispectral imagery registration based on minor and noise component criterion

For multispectral imagery, spatial registration between bands is a very important part of the overall quality of the multispectral imagery product. Due to the significant differences in scene reflectance at different wavelengths, mostly multispectral imagery registration methods are unreliable. In this paper, a robust multispectral imagery registration method is presented. As we know, the spectral information (endmembers) of some pixels will been confused when multispectral imagery is mis-registered. The change of confused endmembers can be estimated through observing minor eigenvalues. Based on this property, a minor and noise component criterion is defined. The best alignment is reached when their minor and noise component is at its minimum. Experiments were conducted using multispectral imagery from the ETM Satellite. And corresponding registration performance curve is given. Multispectral imagery is pre-registered with the method of normalized mutual information. Computer simulations show that the normalized mutual information method may have a deviation about 10 pixels between a pair of images. The method we presented can have a deviation less than 1/4 pixel. Registering curve show that this method is efficient and robust, it can be used for precise multispectral imagery registration.