SAR Image Unsupervised Segmentation Based on A Modified Fuzzy C-means Algorithm

Image segmentation approaches have contributed a lot to help people interpret Synthetic aperture radar (SAR) images well. Fuzzy C-means (FCM), as a clustering method, is the most popular one of SAR image segmentation. However, classical FCM is quite sensitive to speckle noises included in the SAR images, and adding the spectral information is still not enough to improve the performance. Hence, there is a need to take advantage of other useful information to improve segmentation results. In this paper, first of all, Gabor filter is adopted to extract texture features. And then, the SAR image is divided into several patches as the unit to process. At last, FCM with improved nonlocal spatial information (FCM_INLS) is modified with Gaussian kernel distance instead of the Euclidean distance, called FCM_KWINLS, and it can be more robust to speckle noises. In our experiments, the results demonstrate that the proposed method effectively improves the performance of SAR image segmentation.