Research Of Unsupervised Change Detection In Remote Sensing Image

Change detection in remote sensing images of the same scene taken at different times is one of the most interesting topics of image processing. It finds important applications within different fields. Two main approaches, supervised and unsupervised, are used to detect the change. The former requires the availability of a ground truth. The latter performs change detection without any additional information besides the raw images considered. The effective unsupervised change-detection method is fundamental in many applications in that the suitable ground-truth information is not always available.The change detection process performed by unsupervised techniques is usually divided into three main sequential steps: pre-processing, image comparison (to generate difference image) and analysis of the difference image. This paper researches the analysis of the difference image mainly. First, image change detection algorithms used widely are introduced in detail. Then, carrying out experiments on two synthetic data sets and a real data set, the change detection processes of the pixel-dependent unsupervised algorithm and the region-dependent unsupervised algorithm are explained respectively.Usually, gray information is taken into account only in the pixel-dependent unsupervised algorithm, ignoring the characteristics of neighboring pixels. The algorithm is simple and fast, but the detection result isn't ideal. So the algorithm is popularly used in Real-time system, which doesn't require strict precision. And the region-dependent unsupervised algorithms are used in most precise detection applications. Both gray information and adjacent pixels characteristic are taken into account in the algorithm, so the detection precision and anti-noise capability can be improved. But the detection speed is quite low. And edge pixels and strong noise are ignored entirely, or segmented to change type or unchanged type randomly, so that the detection precision is sure to be affected. Because of the shortage of traditional unsupervised algorithm, three algorithms are proposed in the paper based on the spatial information of the difference image and the clustering characteristic of 2-D histogram formed by pixel gray levels and the local average gray levels. The proposed algorithms segment the pixels of...