Research On Feature Extraction And Terrain Classification Of Domestic Full Polarization SAR

Synthetic aperture radar(SAR)is an all-day,all-weather remote sensing information acquisition equipments.Compared with single-polarization SAR,full-polarization SAR(PolSAR)can obtain richer target polarization information,so it has been widely used in the field of feature classification.In recent years,China's domestic satellite-based and airborne full-polarization SAR has been put into operation and experimental data have been acquired in some areas,and the feature classification based on domestic polarization SAR data will play an important role in the field of mapping and land resource survey in China.In order to test the effectiveness of domestic polarized SAR data in feature classification,this paper carries out the research on feature classification based on domestic satellite-based and airborne fully polarized SAR data,and the main research contents of this paper are as follows.Firstly,the polarization features of domestic fully polarized SAR data are extracted by using classical target polarization decomposition methods such as Freeman,Yamaguchi,VanZyl,etc.Through comparison,it is found that different polarization decomposition methods can obtain distinctive features of different types of targets,but a single polarization decomposition method cannot obtain distinctive features of all types of targets.Secondly,classical unsupervised classification such as H-?,wishart-H-? and classical supervised classification such as support vector machine were tried to process the domestic fully polarized data,and the examples were verified and the results were analyzed through the data of White Lake Farm respectively,but the results were not satisfactory and there were more obvious misclassification or omission.Again,considering the principle that the more the amount of heterogeneous features of the input classifier,the higher the classification accuracy,the polarization features obtained by multiple polarization decomposition methods are input to the same classifier,but this will lead to an excessive computational burden of classification training.In this paper,two techniques,wavelet transform(WT)and principal components analysis(PCA),are used to achieve efficient fusion of polarized feature images,which not only suppresses image noise,but also reduces the dimensionality of the classification input set while obtaining effective polarization information,alleviating the The computational burden of Markov clustering segmentation process is reduced.Finally,in this paper,deep learning is introduced into the domestic fully polarized SAR feature classification.In order to further eliminate the duplicated and redundant features obtained from different polarization decompositions,an improved wavelet transform is proposed to reconstruct all the polarization information under the same decomposition mechanism to obtain a feature map in order to suppress noise and reduce the feature dimensionality.Then,it is processed by PCA method to further reduce the feature dimensionality,and finally input to the deep learning Deeplabv3+ network framework for training to complete the target classification,which achieves better classification accuracy while reducing the training computation.In summary,this paper extracts different feature information through target polarization decomposition based on domestic full-polarization remote sensing image data,and uses different methods for feature target classification,and analyzes the accuracy of classification results to verify the effectiveness of domestic full-polarization SAR data in feature classification.