The current synthetic aperture radar (SAR) image change detection still faces the following two challenges. (1) Robustness of difference image (DI) generation. The existing DIs are blurred, and there are more interfering pixels with the same gray value as the change pixels in the DIs. The change regions are influenced by the background information. (2) Effectiveness of DI analysis. In recent years, DI-based unsupervised machine learning or deep learning methods for change detection usually use image sample blocks for spatial feature extraction of pixels to be classified, which lose the detailed information characterizing change information, and there are many false alarms in the classification results, which is not efficient. If we can make full use of various features, reduce false alarms, and improve efficiency, the performance of detection will be greatly improved. Therefore, both robust DIs and diversity feature extraction should be considered to build a robust and fast SAR change detection model. Therefore, this study proposes a new unsupervised change detection method based on the DI of the log-hyperbolic cosine ratio (LHCR) and multi-region feature convolution extreme learning machine (MRFCELM), namely, LHCR_MRFCELM, to solve the problems of poor quality, low detection accuracy, and long detection time in SAR image change detection.

In this study, four methods are experimentally compared and analyzed with the proposed method on four datasets to demonstrate the performance of LHCR_MRFCELM. Figure 5 shows the images of the final detection results of the five methods. Except for the method proposed in this study, all the detection results have many white false alarms. The Kappa value of the LHCR_MRFCELM method (86.44%) as shown in Table 2 is significantly better than that of the rest comparison methods, and the method also takes very little time (20.7 s). In addition, the necessity of each step in the generation process of DIs is discussed. Figure 9 shows the analysis of each step of the DIs generated from the original images, and the situations when one, two, or three steps are missing in the generation process are compared. Among them, the complete DI of LHCR achieves the largest Kappa value in each dataset, which illustrates the necessity of each step of DIs. The neighborhood size r in multi-region feature extraction is also discussed and it is demonstrated that the proposed method has the best performance on four datasets when r is set to 5.

In this study, we propose an unsupervised change detection technique using LHCR_MRFCELM. The method uses the speckle reducing anisotropic diffusion filter, LHCR, and median filtering to generate robust DIs and a fast and efficient MRFCELM to improve the accuracy and efficiency of classification. The log-hyperbolic cosine transformation is a contrast enhancement function used to enhance the contrast between change region and background region in an image, especially the edges where the contrast is not obvious. After DI generation, the HFCM results generated from the DI are used as labels to select sample blocks from the dual-temporal SAR images and the DI, and an MRFCELM is designed to automatically perform feature extraction and classification. Experiments validate the effectiveness of the method, which has better performance than unsupervised change detection methods such as NR_ELM, GaborPCANet, CWNN, and DDNet. The proposed method has no complicated feature extraction steps and no excessive parameter settings, which is easy to use, fast, and stable and has potential for engineering applications.