SAR Image Change Detection Based On Unsupervised Methods

Synthetic Aperture Radar( SAR) which has features of high resolution, all-time and all-weather-work, has been widely used in scientific research, industrial and agricultural production as well as military and other fields. Currently, with the nature environment increasingly changing, urban development is very rapid and all kinds of natural disasters occur frequently. SAR images of all the advantages provide technical support and emergency support for rapid response to disasters and changes. Unsupervised method is currently a hot method to solve the problem with a lot of practical significance. In this paper, SAR image change detection based on unsupervised methods is researched. We worked on a variety of methods to improve the existing and proposed our own innovative algorithms. Our research achievements are as follows:1. Unsupervised feature learning methods have been studied. Combining unsupervised feature learning method and SAR image change detection problem, this part proposes a novel SAR image change detection method based on deep learning. The learning algorithm for deep architectures includes unsupervised feature learning and supervised fine-tuning to complete classification. The unsupervised feature learning aims at learning the representation of the relationships between the two images. And the supervised fine-tuning aims at learning the concepts of the changed and unchanged pixels. The approach accomplishes the detection of the changed and unchanged areas by designing a deep neural network. The main guideline is learning to arrive at a change detection map directly from two images with the trained deep neural network, which omits the process of generating a difference image which shows difference degrees between multi-temporal SAR images, thus avoiding the effect of a difference image on the change detection results. The results on different datasets show that the application of deep learning to change detection obtains satisfactory results.2. SAR image change detection has been studied. The improving strategies are proposed respectively for the construction and analysis of difference map. It is based on an improved neighborhood-ratio operator and modified Kittler and Illingworth thresholding algorithm using generalized Gaussian assumption. An improved neighborhood-ratio technique is proposed to generate a difference image by combining gray level information and spatial information of neighborhood pixels. In order to reduce the impact of the noise further, an improved NR approach applying a weighted operator(WNR) gives each neighborhood pixel a reasonable weight value according to its respective similarity with the center pixel. A modified Kittler-Illingworth threshold selection criterion under the generalized Gaussian(GG_KI) assumption is proposed for modeling the distribution of changed and unchanged classes.3. The analysis of difference map has been studied. An algorithm based on multi-scale with lever-set method is proposed. An energy function known as the piecewise constant approximation Mumford-Shah segmentation model is defined to discriminate the changed and unchanged classes in the difference image. The minimization of this energy functional is realized according to an attractive lever-set method seeking to find an optimal contour which splits the image into two mutually exclusive regions associated with changed and unchanged classes, respectively. In order to achieve a more robust initial contour, stationary Wavelet Transform is used to produce different scales images. Noise and some false edges can be decreased to a certain extent on the coarse scales. Therefore, lever-set evolution on the coarse scales is more robust to noise and avoids the closed contour to stop local minimum or false edges. The result on the coarse scale is taken as the initial contour of the evolution of the lever-set on the finer scale.