Research On Unsupervised Classification Of Polarimetric SAR Images Based On Similarity Network Fusion

Unsupervised polarimetric synthetic aperture radar(SAR)images classification is an essential step in polarimetric SAR images automatic interpretation.However,the studies on unsupervised polarimetric SAR images classification are relatively few,and a large number of problems and challenges still exist in unsupervised polarimetric SAR images classification.Firstly,polarimetric SAR images are seriously corroded by coherent speckle noise,leading to low classification accuracy.Secondly,the number of terrain classes affects the final unsupervised classification results greatly,and in the majority of unsupervised classification algorithms for polarimetric SAR images,the number of terrain classes is given in advance by manual specification according to some prior knowledge.However,the estimation of the number of terrain classes has been studied rarely.Thirdly,many types of feature vectors are usually extracted and stacked directly to a high-dimensional feature vector,thus leading to loss of some features' discriminability.Multi-view learning algorithms can effectively combine the data sets from different views,so that the discriminability of the multiple view data sets becomes stronger.Therefore,each feature vector extracted from a polarimetric SAR image is regarded as a different view data set,which allows multi-view learning algorithms to be introduced into the polarimetric SAR images classification.Similarity Network Fusion(SNF),which is one of multi-view learning methods,can integrate a fused similarity matrix,incorporating the advantages of multiple view data sets,by a cross-network diffusion process.Compared with the similarity matrix formed by each individual view data set,the fused similarity matrix is much more discriminative for classification.Therefore,the research of unsupervised classification of polarimetric SAR images is based on SNF in this paper.The main work and innovation are as follows:1.As for superpixel-based unsupervised classification methods,a superpixel is considered as a processing unit.Therefore,these methods can overcome the effect of speckle noise in polarimetric SAR images as well as with high computation efficiency.Therefore,the research of unsupervised classification of polarimetric SAR is based on superpixels.The first step of superpixel-based classification methods is to oversegment the polarimetric SAR image into superpixels.Most superpixel segmentation algorithms are proposed for optical images.Since the polarimetric SAR images are affected by speckle noise and there are many small or slim regions in them,it is difficult to obtain the ideal result when these methods are applied directly for polarimetric SAR images.In order to solve the above problems,a fast superpixel segmentation algorithm for polarimetric SAR images is proposed in this paper.Firstly,the unstable pixel set and superpixel models are initialized.Secondly,the unstable pixels are relabeled locally based on the fast revised Wishart distance.Thirdly,unstable pixel set and superpixel models are updated.Fourthly,a post-processing procedure based on dissimilarity is used.Finally,extensive experiments based on a simulated image and two real-world polarimetric SAR images are conducted,showing that the proposed algorithm,compared with three state-of-the-art methods,performs better in terms of several common evaluation criteria,with high computational efficiency,accurate boundary adherence and homogeneous regularity.2.The number of terrain classes affects the final classification results greatly,and in the majority of unsupervised classification algorithms for polarimetric SAR images,the number of terrain classes is given in advance by manual specification according to some prior knowledge.However,the estimation of the number of terrain classes in polarimetric SAR images has been studied rarely.To address this problem,a method,which is utilized to estimate the number of terrain classes in polarimetric SAR images,is proposed in this paper.Firstly,the fast superpixel segmentation algorithm for polarimetric SAR images above is utilized to over-segment the polarimetric SAR images.Secondly,a dissimilarity matrix is constructed based on the dissimilarity measure and superpixels.Thirdly,iVATdt(the improved VATdt algorithm by us)is used to estimate the number of terrain classes in the polarimetric SAR images by computer.Finally,Extensive experiments based on one simulated and two real-world polarimetric SAR images verify that this method can effectively estimate the number of terrain classes.3.Numerous feature vectors can be extracted from a polarimetric SAR image.Generally,these feature vectors are directly concatenated into a high-dimensional feature vector for classification of polarimetric SAR images,thus leading to loss of some features' discriminability.To solve this problem,each feature vector extracted from a polarimetric SAR image is regarded as a different view data set,which allows multi-view learning algorithms to be introduced into the polarimetric SAR images classification.Firstly,the polarimetric SAR image is over-segmented to obtain a number of superpixels,and five similarity matrices are respectively constructed from five feature vectors extracted from the polarimetric SAR image based on superpixels.Secondly,Consensus Similarity Network Fusion,which belongs to the multi-view learning algorithms,is used to generate a fused similarity matrix.Thirdly,spectral clustering is performed on the fused similarity matrix.Finally,a novel classification post-processing strategy is proposed to correct the misclassified superpixels.Extensive experimental results conducted on a simulated and two real-world polarimetric SAR images demonstrate the superiority of the proposed method,compared with five other classical methods.