A novel technique for unsupervised change detection in multitemporal SAR Images

The detection of changes that occur on the Earth surface by using multitemporal remote sensing images is one of the most important applications of the remote sensing technology. This depends on the fact that the knowledge of the dynamics of either natural resources or man-made structures is a valuable source of information in decision-making. In this context, optical remote sensing sensors have been used for addressing change-detection applications for many years. Unlike the optical sensors, images acquired by synthetic aperture radar (SAR) have been less exploited in the context of change detection. This is explained by the fact that SAR images suffer from the presence of the speckle noise that renders their analysis complex. However, the use of SAR sensors in change detection is attractive from the operational view-point, since they present the advantage to be independent on atmospheric and sunlight conditions. In the context of multitemporal SAR image analysis, the problem of change detection has been addressed with focus on different aspects, which include the choice of the comparison operator, the image despeckling and the optimal threshold selection. Despite some interesting works have been proposed in the literature, the main problem still open with SAR data is the lack of accurate and reliable methods capable to perform unsupervised change detection in a completely automatic way. In this paper, we propose to face the aforementioned issue by developing an automatic and unsupervised change-detection method specifically oriented to the analysis of multitemporal single-channel single-polarization SAR images. Such a method is based on three main steps: 1) controlled pre-processing based on adaptive filtering (despeckling); 2) comparison of a pair of multitemporal images according to a log-ratio operator; 3) automatic analysis of the log-ratio image. The first step aims at reducing the speckle noise in a controlled way in order to maximize the discrimination capability between changed and unchanged classes. In the second step, the two filtered images are compared in order to generate the log-ratio image. Finally, in the third step, changes are identified by analyzing the log-ratio image using the Kittler & Illingworth (K&I) optimal threshold selection algorithm reformulated under the Generalized Gaussian (GG) distribution assumption for the changed and unchanged classes. The choice of this model is attractive, since it is characterized by a small number of parameters and can approximate a large class of statistical distributions (e.g., impulsive, Laplace, Gaussian and uniform distributions). In order to take account the effect of the filtering process on the accuracy of the change-detection results, we propose to identify the optimal number of filtering iterations automatically by means of the minimization of the K&I criterion over the filter iterations instead of using empirical methods (as usually adopted in the literature of SAR). Experimental results carried out on two different real data sets confirm the effectiveness of the proposed approach, which results in higher change-detection accuracies with respect to other methods usually adopted with multitemporal SAR images.