PolSAR Image Classification Based On Complex-valued Neural Network

Deep learning plays a very important role in the field of polararimetric synthetic aperture radar(PolSAR)image classification,whose purpose is to train the model through a large amount of labeled data.Which could mine richer image information and get higher classification accuracy.Traditional neural network extracts image features and classifies image which both are based on real-valued domain.Because of its simplicity and effectiveness,it has been widely used in PolSAR image classification.However,the traditional neural network classification algorithms do not consider the complex characteristics,phase information,spatial information,lack of labeled samples and low credibility of PolSAR data.As a result,PolSAR data cannot be used more effectively,and the ideal classification result cannot be obtained.In view of the above problems,our study not only considers the complex characteristics,phase and spatial information of PolSAR image into complex-valued neural network(CV-NN),but also focus on solving the problems of lacking labeled samples and low credibility.Based on this,three PolSAR image classification algorithms which base on CV-NN are proposed.The specific research work is introduced as follows:(1)By introducing the phase and complex-valued information of PolSAR images,this thesis proposes a complex-valued Wishart stacked auto-encoder network(CV-WSAE)for PolSAR image classification.The algorithm stacks a complex-valued Wishart auto-encoder network(CV-WAE)and a complex-valued auto-encoder network(CV-AE),and then all elements in the forward and back propagation are extended to the complex-valued domain in order to obtain richer image information.Finally,a linear classifier is used to accomplish the classification task of PolSAR image.Experiments show that the CV-WSAE classification algorithm can extract more effective image information and obtain relatively ideal classification results on PolSAR images.(2)In order to compensate for the loss of image information caused by the traditional PolSAR image classification algorithm,and the problem of fewer labeled samples and serious network overfitting,this thesis proposes a semi-supervised recurrent complex convolutional neural network(RCV-CNN)for PolSAR image classification.The algorithm first uses Wishart classifier to select a small number of highly reliable samples;then it is used as the input data of CV-CNN to improve the classification accuracy of the model based on enhancing the spatial information;finally,the semi-supervised recurrent strategy is used to continuously predict the sample to be tested to expand the labeled sample set,thereby achieving better classification performance.In particular,the classification algorithms under different recurrent strategies are named RCV-CNN1 and RCV-CNN2,respectively.Experiments show that the proposed RCV-CNN algorithm achieves higher accuracy and better spatial consistency.In addition,when the number of training samples is small,the RCV-CNN algorithm can solve the problem of network overfitting effectively.(3)In order to improve the credibility of labeled samples and the classification performance of the RCV-CNN classification algorithm,this thesis proposes a semi-supervised complex-valued convolution neural nestwork network with tri-training algorithm(Tri-CV-CNN).First,Wishart classifier is used to select some samples with high reliability;then tri-training algorithm is applied to label some unlabeled samples to increase the reliability of expanded labeled samples and the generalization ability of classification model;finally,the highly reliable extended label sample sets are gotten input into the CV-CNN model to accomplish the process of network training.Experiments show that the Tri-CV-CNN algorithm not only improve the credibility of expanded training samples sets,but also increase the classification accuracy and spatial consistency of model.