Change Detection In SAR Images Using Spectral Clustering

Change detection is an important application of remote sensing image, which is a process aimed at identifying the earth’s surface changes by analyzing images acquired on the same geographical area at different time. Such important applications of change detection include dynamic changing study on land use, burned areas identification, investigation of resources, environment monitoring, military affairs and civil, etc. So far, many scientists have developed many methods for change detection, and these techniques can be divided into two kinds:the supervised approach and the unsupervised approach. All supervised techniques are based on supervised classification methods, which require the availability of a multitemporal ground truth in order to derive a suitable training set for learning process of the classifiers. In unsupervised techniques, there is no need of training data. Although the supervised methods have the advantage of explicitly identifying the earth’s surface changes over the unsupervised techniques, in practice the supervised methods are difficult to generate an appropriate multitemporal ground truth. Accordingly, in many applications in which a ground truth is not available, the unsupervised techniques are more useful than the supervised ones for change detection. Accordingly, in literature, three steps are suggested to unsupervised change detection. They are pre-processing, image comparison and image analysis. In this paper, through the research on and analysis of the existing problems of the traditional change detection methods, we improve these algorithms and propose three new algorithms. The major works can be summarized as follows:(1) An unsupervised change detection method based on spectral clustering and non-local difference image method is proposed. The novel difference image is generated by integrating the typical difference image method with non-local filter, which exploits both the spatial neighborhood information and gray similarity information. It can well reduce the noises. The spectral clustering algorithm can cluster the difference image into two clusters and get the good change map. Compared with traditional clustering algorithms, such as k-means, spectral clustering can recognize the clusters of unusual shapes and obtain the globally optimal solutions.(2) A new approach based on histogram spectral clustering and bilateral difference image methods is proposed. The difference image is generated by employing the spatial neighborhood information, which can well result in enhancing the change information and inhibiting the unchanged information. However, the traditional spectral clustering suffers from the scalability problems in both memory use and computational time. In this work we propose the histogram spectral clustering which can alleviate both memory and computational bottlenecks to perform clustering on larger image datasets.(3) A novel approach based on low-rank matrix decomposition and k-means method is proposed. This algorithms use low-rank to generate the difference image. The k-means algorithm can cluster the difference image into two clusters and get the change map. Compared with traditional change detection algorithms, the difference image based on low-rank can recognize the changed information in images. Experimental results show that the proposed method is suitable for change detection.