Research On PolSAR Classification Based On Deep Learning

As an advanced active earth observation remote sensing system,Polarimetric synthetic aperture radar(PolSAR)emits electromagnetic waves with two different polarizing modes,horizontal and vertical,then receives the amplitude and phase information of the back-scattering echo generated by the interaction with the target to obtain the polarimetric scattering characteristics By adjusting the different combination mode of transmitting and receiving the polarimetric electromagnetic waves,the acquisition ability of ground object information is greatly improved.As an important part of PolSAR technology,polarimetric SAR image classification has wide application and great research value in agriculture,military,city planning,environmental protection and resource development.In recent years,with the outstanding achievements of deep learning in the fields of target recognition,image classification,and automatic driving,more and more researchers have turned their attention to application of deep learning.In the research of remote sensing image classification,deep learning algorithm also shows great potential to surpass traditional image classification methods.In this paper,the deep convolution network is applied to polarimetric SAR image classification.In this paper,the deep convolutional network is applied to the classification of polarimetric SAR images,and the classification method that combines the The polarimetric decomposition which characterizes the scattering characteristics of ground objects with the deep convolution network is studied.Based on the full analysis of the characteristics of polarimetric SAR images,the deep convolutional network is improved to extract the deep features of the polarimetric SAR data and achieve better classification results.The research work of this paper mainly includes the following three aspects:1.Introduce the basic structure and network characteristics of classical semantic segmentation network-Segnet.Analyze the dvantages of deep learning in the field of polarimetric SAR image classification.The classification task is realized by combining polarimetric SAR pseudo color image and Segnet network,and compared with traditional image classification methods.2.The widely used U-net network is introduced in detail,and its principle of making full use of image semantic information is analyzed.Aiming at the problem that polarimetric SAR data information can not be fully utilized in classification tasks,a classification method based on multi-features combined with U-net is proposed,which effectively improves the classification accuracy.3.By analyzing the respective characteristics of Segnet and U-net,we propose a classification method based on improved U-net which introduce the Inception structure that can extract multi-scale feature,and then combined with the post-processing by super-pixel segmentation.The proposed method can better eliminate the outliers of misclassification,maintain the boundary of objects and improve the classification accuracy.