Combining MRF And V-SVM For SAR Sea Ice Image Classification

Sea ice is an important factor for the global climate system. For the transportation and marine resources development work, we must strengthen the monitoring and forecast of sea ice, synthetic aperture radar(SAR) is an important tool of monitoring sea ice.The commonly used method of sea ice images interpretation is sea ice images classification. In this paper, the spatial correlation based on Markov random field (Markov Random Field, MRF) will be introduced support vector machine (Support Vector Machine, SVM) classifier. Due to the spatial context information based on MRF can improve the accuracy of the scene interpretation, while SVM has an high generalization ability, efficient performance of classification and robustness of the Hughes phenomenon, and it can overcome the impact of statistical estimation error, therefore spatial interaction modeling by using MRF in the framework of SVM is feasible. In this paper, we proposed MRF-vSVC classification system by fully integrated with the advantages of MRF and SVM in the field of remote sensing image classification.First,the system will regional SAR sea ice images, get the area of the sample of classification and the information of edge, collect the right amount of training sample sub-image,and extract gray and three texture statistics of the sample area to constituted feature vectors. We get initial classification markings of image by inputting the training data in v-SVC. Then we determine the strength of the edge by proposing dual-threshold criteria. For the area of fuzzy edge, improved model of spatial context (the model of edge context)will be Introduced in v-SVM. By correcting the deviation factor of original problem of v-SVM and solving the optimal solution corresponding dual problem, we ultimately get the classification results.Unlike the past, this method has in the following two different aspects:(1) In order to improve the nonstationarity adaptability of SAR sea ice image, this paper will optimize edge context model by using space context model based on neighborhood, construct the deviation factor based on edge context of v-SVM, so the information of edge context based on MRF will be introduced tov-SVM, which is to modify the optimal hyperplane.(2) To improve the efficiency of the algorithm, this paper set up a dual threshold based on the strength of edge.It retains the label of region for strong edge, removes the edge and edge identification for weak edge, and optimizes classification by considering the edge context information for fuzzy edges,for which grayscale did not change significantly but may be connected to different types.