The multi-spectral optical remote sensing image registration is one of the important steps among the ground station preprocessing chain.The accuracy of registration is crucial to product quality assurance.Generally

an image correlation method based on pixel DN is widely used in the multi-spectral images registration processing.However

the registration accuracy of correlation method depends on the character of the landscape in the space camera’s view.The discrepancies of band-to-band registration are commonly observed in those images which are short of manifest feature points such as large scale water surface

desert

meadow

forest and seaport.In this paper

a linear CCD geometric bias model along CCD array and orbit directions of the BJ-1 small satellite multi-spectral camera is presented.The measured geometry bias is a constant value which is derived from the large amount of tie points between NIR and red or green and red bands along the image row and column directions.Finally

the experiment shows that the matched BJ-1 multi-spectral images can achieve sub-pixel registration accuracy without calculating the pixel DN correlation value between the raw images.And the proposed method has been used in the BJ-1 and DMC

U

K2 small satellite multi-spectral images registration preprocessing.