Study On The Application Of Nonnegative Matrix Decomposition In Remote Sensing Image Change Detection

Remote sensing image change detection is an application of remote sensing technology which acquires the change information of images via comparing and analyzing two or more remote sensing images for the same area in different periods. It has been widely used in many fields such as economic and national defense construction.Nonnegative matrix decomposition algorithm is a method of matrix decomposition newly proposed. It is a kind of very important matrix dimension reduction techniques. The application field of NMF is very broad, such as image processing, computer vision, text analysis and so on. This paper tries to apply NMF to remote sensing image change detection. It mainly includes the following three contents:(1) Using NMF to fuse two difference images. Conventional change detection usually takes a single difference image as the research object, but a single difference image often has limitations. Aimed at this problem, this paper proposes a method of change detection which uses NMF to do fusion. To SAR image, we fuse its logarithmic ratio image with MRD image. To optical image, we fuse its difference image with t-test image. We acquire a good effect through simulation experiment. And we verify the effectiveness of the proposed method via compared with the existing several change detection methods.(2) Change detection based on the focus area. In order to reduce false alarm rate caused by the noise, and miss alarm rate caused by weakness of the change information, this paper adopts the method of confirming the focus area for change detection. Firstly, we produce texture image of difference image using gray level co-occurrence matrix. We use variance texture map to foreshadow confirming the focus area because variance can highlight change area boundary and separability is stronger. We utilize NMF to extract texture feature of the background. And we weaken the texture feature in background region through calculating Euclidean distance between neighborhood image block of every pixel of texture image and the texture feature image of background. Then we multiply corresponding pixels of the original difference image and the proposed image. So, we can get a ideal change profile image. We cluster it and fill the inner region through inflation in order to obtain the focus area. In the end, we modify the original difference image according to the focus area and we acquire a good change detection result through dealing with the modified difference image.(3) We research the characteristics of NMF in clustering. Considering the problem that the difference image has noise, we use bilateral filtering to solve it. Then we acquire the eigenvector of each pixel via scanning the second order neighborhood of it. Moreover, we use PCA to do dimensionality reduction and adopt Semi-NMF to do clustering in the end. We can see that the algorithm has good detection results through comparing with the front algorithms in this paper. And Semi-NMF express high accuracy through comparing with K-means.