Superpixel-Oriented Unsupervised Classification for Polarimetric SAR Images Based on Consensus Similarity Network Fusion

Unsupervised polarimetric synthetic aperture radar (PolSAR) image classification is an important task in PolSAR automatic image analysis and interpretation. Generally, a group of features is insufficient to effectively classify PolSAR images, especially in multiple terrain scenarios. Therefore, multiple features need to be extracted for PolSAR image classification. However, how to combine and integrate these features effectively to fully utilize each feature's information and discriminability need to be determined. Such integrated work has traditionally received little attention. In this paper, a novel unsupervised classification framework for PolSAR images is proposed. First, a PolSAR image is oversegmented via a fast superpixel segmentation method. Second, five feature vectors are extracted from PolSAR images via superpixels, resulting in five corresponding similarity matrices that are constructed by using Gaussian kernels. Third, consensus similarity network fusion (CSNF), originally proposed and widely used for biomedical sciences, is employed to combine and integrate the five similarity matrices to obtain a fused similarity matrix. Fourth, spectral clustering method, based on the fused similarity matrix, is used to cluster the PolSAR image. Finally, a novel classification postprocessing procedure is presented and exploited to smooth the initial clusters and correct some misclassified pixels. The extensive experimental results conducted on one simulated and two real-world PolSAR images demonstrate the feasibility and superiority of the proposed method compared with five other state-of-the-art classification approaches.