Unsupervised Classification Of PolSAR Images Based On Tensor Product Graph Diffusion

Unsupervised polarimetric synthetic aperture radar(PolSAR)image classification is an important task in PolSAR automatic image analysis and interpretation.Concerning the difficulty of similarity expression and the effects of speckle noise in PolSAR unsupervised classification,the research on the unsupervised PolSAR image terrain classification based on Tensor Product Graph(TPG)diffusion is carried out.1.In order to improve the computational efficiency and reduce the influence of speckle noise on the classification results,a fast superpixel segmentation algorithm based on regular hexagonal initialization is proposed.The clustering center is placed in the regular hexagon,compared to square initialization,the number of points is reduced when searching for the center point of the superpixels,and the computational efficiency is remarkably improved.The experimental results show that the proposed method can take into account the segmentation performance and computational efficiency,and meet the requirements of superpixel segmentation of PolSAR images.2.For the sake of effectively improving the classification accuracy,combined with the classification ability of multi-view data,an unsupervised classification method based on cross-view tensor product diffusion is proposed.After extracting multiple features based on superpixels and combining three feature vectors,each feature vector is regarded as one view data,and then a diffusion process is implemented on the constructed cross-view tensor product graph to produce a similarity matrix combined with multiple views.The proposed method is compared with 4 unsupervised classification algorithms based on pixel/superpixel,which verifies the validity of superpixel-based classification method and the strong discrimination ability of the proposed method.3.For better mining the intrinsic similarity information between data points,an unsupervised classification method based on tensor product diffusion is proposed.After extracting multiple features based on superpixels to form the high-dimensional feature vector,the constructed original similarity matrix is used to learn the context similarity information by tensor product diffusion,generating the similarity matrix with stronger discriminative ability and classification ability.Finally,spectral clustering based on the diffused similarity matrix is adopted to achieve the terrain classification results.Extensive experiments conducted on both the simulated and real-world PolSAR images demonstrate that the proposed method can robustly measure the geodesic distance between data points and improve the classification performance.