An Unsupervised Change Detection Algorithm Based on Spectral Signature Analysis in Multispectral Images

In this work an innovative unsupervised change detection technique for multispectral and hyperspectral images is proposed. It consists in the comparison between two co-registered and radiometrically correct multispectral images, carried out by comparing pixels’ spectral signatures. Supposing a change is occurred, we can state that the spectral vector associated to a pixel in the “changed image” makes an angle with the spectral vector of the corresponding pixel in the reference image or has a different norm. Hence in our approach we have used for the classification of changed areas two main features based on the direct or indirect measure of the angle and of the norm of changed vector projection on the corresponding spectral vector in the reference image. More in deep, since cosine and sine functions represent an indirect measure of this angle, we can calculate these values to obtain the changes we are looking for. Nevertheless, measuring an angle that is equal to zero doesn’t mean that no change has occurred, because there might have been just a variation along the reference spectral vector’s direction that has no effect on angle but only on pixel brightness. So we have defined another change indicator, the Brightness Change Factor, whose value gives us the information about changes occurred in pixels’ radiance values.