Research On Polarimetric SAR Image Classification Algorithms

Polarimetric Synthetic Aperture Radar（Pol SAR）,as an active microwave remote sensing imaging system,has the characteristics of all-day and all-weather operation.Compared with traditional Synthetic Aperture Radar（SAR）,it records more complete target backscatter information.Polarimetric SAR image classification,as a key technology in polarimetric SAR image processing and information extraction,directly affects the application effect in the fields of surface mapping,environmental analysis,resource survey,geological distribution,national defense and military,etc.Research on polarimetric SAR image classification algorithms is of great importance.theoretical significance and practical value.In this paper,based on the polarization characteristics and scattering characteristics of polarimetric SAR images,the research on unsupervised classification algorithms of polarimetric SAR images is carried out.Based on Cloude-Pottier decomposition and Freeman decomposition respectively,initial classification is performed according to the target decomposition results.The H/α/ A Wishart classification algorithm based on complex Wishart distribution and maximum likelihood estimation algorithm is introduced,and through multiple clustering iterations,the boundary between classes in the initial classification result is further optimized,and the classification accuracy is improved.Aiming at the problem of uneven initialization classification,this paper proposes an improved initialization classification method combining Cloude-Pottier decomposition and Freeman decomposition.The improved method uses the polarization entropy and average scattering angle of the Cloude-Pottier decomposition results,and the scattering characteristics of the Freeman decomposition results to redefine the initialization classification rules.The original polarimetric SAR image is segmented.Compared with the traditional unsupervised classification algorithm,it effectively reduces the misclassification in the initialization classification,and improves the classification efficiency and accuracy.The overall accuracy is increased from 89.0% of the H/αWishart classification algorithm to 93.2% of the improved algorithm.The experimental results and ground-truth data with a high degree of consistency.The Convolutional Neural Networks（CNN）algorithm in deep learning is introduced into polarimetric SAR image processing,and the semantic segmentation of polarimetric SAR images is realized by using the U-Net network model.Aiming at the limited polarimetric SAR data and the lack of real object labels,data enhancement algorithms are used to expand the data set,and the diversity of the data set is improved by rotating,flipping,cropping and other methods.In the traditional optical image processing method,each polarimetric SAR image is processed as a single-channel grayscale image,completely ignoring the relationship between different polarization methods,and it is difficult to obtain a better classification result.On this basis,the polarimetric feature of polarimetric SAR images is introduced,and the single channel is improved to fourchannel input containing polarimetric information,and the four polarimetric SAR images correspond to a real object annotation data.After comparing several sets of experimental data,the accuracy of the algorithm is increased to about 90%,and the Kappa coefficient is increased to more than 0.8,which verifies the effectiveness and superiority of the algorithm.