For multitemporal synthetic aperture radar (SAR) images, the change detection methods based on nonlocal theory can well suppress the adverse effects of coherent speckle noise over the change detection results. However, effectively retaining the edge information of the changed area is still a challenging task. To overcome this problem, this study proposes a change detection method based on progressive nonlocal theory. First, the progressive nonlocal theory is used to extract the spatial–temporal nonlocal information from multitemporal SAR images. Compared with the traditional nonlocal theory, the progressive nonlocal theory proposed in this study has three distinctive characteristics: 1) the progressive nonlocal neighborhood from the matching window to the search window; 2) the progressive optimization of matching window weight from the isotropic Gaussian distribution to the irregular distribution; and 3) the progressive increase of noise level from the 2 sigma principle to the 4/3 sigma principle (the noise level corresponding to the 4/3 sigma principle is 1.5 times the noise level corresponding to the 2 sigma principle). The difference image is then obtained by using the spatial–temporal nonlocal information and the ratio operator. Finally, the change map is obtained by applying a threshold segmentation method to the difference image. Two data sets were used for the testing, and it was shown that compared with other advanced methods, the method proposed in this study can better retain the edge information of the changed area and improve the Kappa coefficient and F1 score of the change map.

中文翻译：

---

基于渐进非局部理论的SAR图像变化检测


对于多时相合成孔径雷达(SAR)图像，基于非局部理论的变化检测方法可以很好地抑制相干散斑噪声对变化检测结果的不利影响。然而，有效保留变化区域的边缘信息仍然是一项具有挑战性的任务。为了克服这个问题，本研究提出了一种基于渐进非局部理论的变化检测方法。首先，采用渐进非局域理论从多时相SAR图像中提取时空非局域信息。与传统非局部理论相比，本研究提出的渐进非局部理论具有三个显着特征：1）从匹配窗口到搜索窗口的渐进非局部邻域； 2）匹配窗口权值从各向同性高斯分布向不规则分布逐步优化； 3)噪声水平从2 sigma原则逐步增加到4/3 sigma原则(4/3 sigma原则对应的噪声水平是2 sigma原则对应的噪声水平的1.5倍)。然后利用时空非局部信息和比率算子获得差异图像。最后，对差异图像应用阈值分割方法得到变化图。使用两个数据集进行测试，结果表明，与其他先进方法相比，本研究提出的方法能够更好地保留变化区域的边缘信息，提高变化图的Kappa系数和F1分数。