PolSAR Classification Based On Robust Kernel Fuzzy C-means Clustering

Polarimetric Synthetic Aperture Radar(PolSAR)is a high-resolution radar,effective image interpretation technology can reflect the real value of PolSAR.Synthetic Aperture Radar(SAR)image classification is a complex and key step in SAR image interpretation.In classification,unsupervised classification can save a lot of human resources compared with supervised classification,and how to improve the classification accuracy and better realize image interpretation is of great significance.At present,many classification algorithms have been proposed for the PolSAR image.In view of the problem that the unsupervised classification method PolSAR images strengthens the discrimination ability and avoids the decrease of classification accuracy caused by noise,a classification method of PolSAR images based on Robust Kernel Fuzzy C-means Clustering(RKFCM)is proposed.The main research contents are as follows:(1)The characteristic parameters of PolSAR image are extracted.By referring to the H/?surface and comparing and analyzing the difference of scattering characteristics of objects around the polarimetric characteristic information,three polarimetric feature parameters,Shannon entropy(SE),radar vegetation index(RVI)and total polarization power(SPAN),are selected to combine to form a new polarimetric feature space,which strengthens the discrimination ability of different objects.(2)The classical FCM(Fuzzy C-means)algorithm is optimized by using the mean-standard method instead of the random method to select the cluster center to avoid falling into local optimum.Aiming at the influence of noise and outliers on the objective function,the clustering center is used to represent the behavior function of the system to restrict the objective function,and the Gaussian kernel function is introduced as the distance measure to replace the Euclidean distance of the linear metric function,and the robust kernel FCM algorithm is obtained.By comparing the processing results of synthetic gray image and optical image with noise,it is verified that the algorithm has better inhibitory effect on outliers and noise in pixels.(3)The SE/RVI/SPAN polarimetric feature space is composed of feature parameters extracted from PolSAR data,and an unsupervised classification method for PolSAR image is proposed by combining the optimized RKFCM algorithm.Not only the main scattering features are retained,but also the outliers and speckle noise in pixels are well suppressed.By using GF-3 PolSAR data for experiments,The proposed method is compared with the classical Freeman-Wishart,H/?-FCM and Pauli-MRF(Madras Rubber Factory)classification methods,the experiments show that the proposed PolSAR classification method based on RKFCM algorithm is feasible and effective,and the overall classification accuracy reaches88.9%,which has higher classi-fication performance than other algorithms.This paper includes 22 figures,5 tables and 74 references.