A change detection approach of high-resolution imagery combined the pre-classification with the post-classification comparison

This paper presents a change detection approach for high-resolution remote sensing image combined the pre-classification with the post-classification comparison. Firstly, a spectrum correction image is constructed on the basis of fusing the differential and ratio algorithm, which can alleviate the effects of single band's sensitivity. Secondly, a binary change mask is extracted from the spectrum correction image by using Otsu. Then multiply it with the overlaying multi-temporal image to extract reliable change region. Finally, the transference process is realized by using an improved Fuzzy C-Means algorithm. Additionally, considering that Fuzzy C-Means cannot determine the cluster center and cluster number automatically, and also be sensitive to the noise, the article utilizes membership entropy to determine the optimal cluster number, and adopts the max-minimum distance to initialize the cluster center, which can avoid trapping in local optimum caused by selecting cluster center randomly, and also introduces the similarity-weight of the neighbor pixels to correct the membership function. Experiments show that the combination of pre-classification with post-classification comparison can realize the whole change process and also alleviate error propagation effects of multi-classification, which makes the change detection results closer to the real situation. Additionally, the improved Fuzzy C-Means has a better performance in determining the cluster number and center as well as reducing the noise interference.