Polarimetric SAR Image Classification Via Deep Learning

With the increasing maturity of coherent microwave remote sensing technology,polarimetric synthetic aperture radar(Pol SAR)is able to image in adverse environments such as nighttime,clouds,and fog and to observe the objects at all day,and has the ability to observe the earth at all times and all days.So,it's widely used in urban planning,land monitoring.In these applications,the classification of Pol SAR images with complex surface features has always been one of the most important key technologies,so it has been a hot and difficult point of research.The fast rise of deep learning has changed the dilemma of traditional algorithms.Especially in the field of image processing,deep learning has shown absolute advantages.In particular,Convolutional Neural Networks(CNN)are widely used in image classification for the ability to mine the deep features of images autonomously.Generative adversarial networks(GANs)optimize the generator and the discriminator simultaneously under the guidance of game theory,which receive widespread attention and application.Despite this,the effect of speckle noise cannot be eliminated.Super-pixels,as an effective method for over-segmenting,can alleviate speckle noise while retaining useful information of images.In this paper,deep learning is combined with super-pixels to obtain robust Pol SAR image classification results.The main tasks are as follows:(1)Aiming at the problem of incomplete target boundary preservation,the superpixel generation of Pol SAR images using deep learning is studied.Based on the synthesized polarimetric pseudo-color image,a pixel affinity network is used to learn superpixels.The pixel affinity network can mine deep polarization features,obtain vertical and horizontal affinity maps of ground objects,and generate superpixels based on these affinity maps.The experimental results show that the superpixel segmentation obtained by the proposed method is better than the existing methods such as SLIC.(2)Aiming at the problem that the interaction relationship between color channels cannot be learned,a classification method of the quaternion CNN combined with superpixels for Pol SAR images is studied.Unlike CNN,which learns the features of each color channel separately,the quaternion CNN combines the three color channels into a quaternion matrix.Based on the quaternion matrix,the convolution kernel is used to mine the deep features and obtain the classification result.Then the idea of nearest neighbors is utilized to optimize the classification results with superpixels.Experimental results show that the classification performance of quaternion CNN is better than CNN,and the classification results combined with superpixel is better than the existing deep learning methods.(3)Aiming at the problem of insufficient classification features,a CNN network GAN-assisted is proposed.GANs have the powerful abinity to learn deep features,especially the intermediate features learned by the discriminator.Embedding these intermediate features into the intermediate features of the CNN network can improve the classification performance of CNNs.Experimental results show that GAN-assisted CNN classification performance is better than CNN.In addition,the classification results combined with superpixels will be superior to existing deep learning methods.