A Novel Method of Unsupervised Change Detection Using Multi-Temporal PolSAR Images
The existing unsupervised change detection methods using full-polarimetric synthetic aperture radar (PolSAR) do not use all the polarimetric information, and the results are subject to the influence of noise. In order to solve these problems, a novel automatic and unsupervised change detection approach based on multi-temporal full PolSAR images is presented in this paper. The proposed method integrates the advantages of the test statistic, generalized statistical region merging (GSRM), and
... ng (GSRM), and generalized Gaussian mixture model (GMM) techniques. It involves three main steps: (1) the difference image (DI) is obtained by the likelihood-ratio parameter based on a test statistic; (2) the GSRM method is applied to the DI; and (3) the DI, after segmentation, is automatically analyzed by the generalized GMM to generate the change detection map. The generalized GMM is derived under a non-Gaussian assumption for modeling the distributions of the changed and unchanged classes, and automatically identifies the optimal number of components. The efficiency of the proposed method is demonstrated with multi-temporal PolSAR images acquired by Radarsat-2 over the city of Wuhan in China. The experimental results show that the overall accuracy of the change detection results is improved and the false alarm rate reduced, when compared with some of the traditional change detection methods. Most of the supervised change detection approaches are dependent on the classification accuracy of the multi-temporal PolSAR images, and are liable to be affected by the significant cumulative error caused by the single remote sensing image classification during change detection [16, 17] . Because of the impact of speckle noise, it is difficult to obtain high-precision classification results from multi-temporal PolSAR images, which limits the use of change detection based on PolSAR images. Unlike supervised change detection approaches, the unsupervised change detection methods implement change detection by directly comparing images acquired at different times, and these methods are widely used in change detection based on multi-temporal PolSAR images  . The traditional unsupervised methods based on PolSAR images usually take the pixel as the basic processing unit. The process of the state-of-the-art unsupervised change detection methods includes: (1) preprocessing; (2) generation of the difference image (DI); and (3) making a decision (thresholding algorithm or clustering algorithm) based on the analysis of the DI  . The preprocessing of multi-temporal PolSAR images mainly consists of radiometric calibration, speckle filtering, and image co-registration, which are all critical to the change detection. In particular, speckle filtering is usually carried out to suppress the speckle before the change detection of PolSAR images. Despite preprocessing, the result is still subject to the influence of noise and false alarms, which are caused by the pixel-based change detection method  . In generating the DI step, two preprocessed images of the same geographical area at different times are compared to generate the DI. There are many methods used to generate the DI using two co-registered images, including the ratio or log-ratio operator of SAR amplitudes or intensities [11, 13, 14] , the hidden Markov chain model [16, 17] , and the Kullback-Leibler divergence method  . These methods are usually applied in multi-temporal single-channel SAR change detection. Unlike the above methods, test statistics can be applied not only to single-channel SAR data, but also to full PolSAR images  . Using multichannel PolSAR data (coherency C 3 or covariance matrix T 3 ) can obtain a more accurate DI  . In the step of change detection analysis, the change detection map can be obtained by a thresholding algorithm or a clustering algorithm. A number of different algorithms can be used to automatically make a decision, such as the k-means algorithm  , the fuzzy c-means (FCM) algorithm , Otsu's thresholding method , Kapur's entropy algorithm  , the Kittler and Illingworth (K&I) algorithm  , and two-dimensional entropic segmentation (TDES)  . However, most of the methods of making a decision are based on the condition of a Gaussian assumption for the probability density function (PDF) of the DI for modeling the distributions of the changed and unchanged classes. Fortunately, the Gaussian mixture model (GMM) is capable of better fitting the arbitrarily conditional densities of the classes in the DI [27,28], but it is still difficult to select the optimal number of components for the GMM . Above all, the existing pixel-based unsupervised change detection methods using full PolSAR images all show certain deficiencies, such as not taking advantage of the full-polarimetric information and being subject to the influence of noise and false alarms  . Meanwhile, the existing object-based unsupervised change detection methods using PolSAR images can suppress the influence of noise and improve the overall accuracy (OA) of the change detection results , but they have difficulty in capturing the global property of the image. The results are also sensitive to segmentation, and this is difficult to accomplish due to the degradation of spatial details and fine structures  . To solve the above problems, a novel method of unsupervised change detection is proposed in this paper, integrating the respective advantages of the test statistic, generalized statistical region merging (GSRM), and generalized GMM techniques. The use of a test statistic is a good strategy for obtaining the DI from multi-temporal PolSAR images. We use the GSRM algorithm to separate the same parts of the DI, which helps us to choose the threshold by generalized GMM. This paper is organized into five sections. In Section 2, the proposed change detection framework is described, and the methods of test statistic, GSRM, and generalized GMM techniques are introduced. Section 3 details the results of the proposed approach on multi-temporal PolSAR images from the city of Wuhan, China. Section 4 discusses the results of the case study. Finally, the conclusions are drawn in Section 5.