Research On Automatic Segmentation Algorithm Of SAR Sea Ice Image Based On MRF

As the number of human activities in the polar regions increases and the global temperature rises year by year,the demand for sea ice testing will increase.Because the Synthetic Aperture Radar?SAR?system has unique advantages such as strong cloud penetration,day/night operation,and wide coverage,thousands of SAR sea ice images need to be processed each year.Support its business activities^[1][2],such as ship navigation and environmental monitoring.For this reason,the rapid processing and interpretation of large-capacity satellite data is urgently needed.Since the acquired SAR data is large,manual interpretation is very time consuming,so it is hoped that it can be automatically interpreted by a computer.At present,in the process of automatically segmenting SAR sea ice images based on unsupervised algorithms,it is still necessary for the interpreter to manually input the number of categories according to prior knowledge,and then automatically interpret the SAR sea ice images.On the other hand,the initial clustering centers in the commonly used GMMs are random,resulting in unstable clustering results,too many iterations,and long running times.In this paper,for the first problem above,we propose to add the Goodness of Fit?GOF?to the existing MRF algorithm to determine the number of image categories.Therefore,instead of manually inputting the number of categories by experience,the purpose of automatic analysis of the image is achieved.For the initial point random selection problem in GMM,we automatically determine the number of ice classes in K-means via the merge operation to provide the initial cluster midpoint for the GMM algorithm.The the segmentation precision is improved,and the segmentation time is reduced.Compared with previous research methods,the research methods in this paper mainly have the following two aspects:1.In order to improve the segmentation efficiency and segmentation accuracy of the algorithm,an improved segmentation method based on GMM model and K-means is proposed.The algorithm integrates the merge operation in K-means clustering,provides the initial value for the GMM algorithm,replaces the random selection process of the initial point in the previous GMM algorithm,reduces the number of iterations of the parameter optimization process,and reduces the complexity of the algorithm.Tests have shown that the segmentation accuracy can be improved and the segmentation time can be reduced to some extent.2.In order to solve the problem that the number of categories needs to be manually input in the segmentation algorithm,based on the feature data of the SAR image category approximating Gaussian distribution or unimodal distribution,the best fitting test is added to the existing MRF algorithm to judge the GMM.The degree of fitting of the probability density function to the image,at a certain confidence level,determines the number of image categories by analyzing the significant class weights of the GMM model,and outputs the feature model parameters at this time to the segmentation algorithm.Thereby achieving the purpose of automatic analysis of the image.