Polarimetric SAR Image Classification Based On Deep Adversarial Learning

Nowadays,Polarimetric Synthetic Aperture Radar(PolSAR)has been widely used in many fields.PolSAR has multiple polarimetric channels,which can obtain more abundant ground object information.As a very important part in the interpretation of PolSAR data,the research work of PolSAR image classification has attracted the attention of scientific researchers and scholars.Deep learning is a new branch of machine learning.It senses important parts of data by simulating human brain structure,learns more abstract high-level information from low-level features,and efficiently and autonomously extracts the internal characteristics of data.Each layer in the deep neural network represents a different level of abstraction.In recent years,with the development of deep learning,a large number of excellent deep learning models have been applied to the classification of PolSAR images.Considering that the characteristics of PolSAR images are complex and large-scale,three methods are proposed to realize terrain classification of PolSAR based on deep adversarial learning algorithm:1.A PolSAR image classification algorithm based on Adversarial AutoEncoder(AAE)is proposed.In this method,the target decomposition features and image texture features of PolSAR image are combined to form the polarimetric-texture combination features.Then,feature extraction and terrain classification are realized after the training of autoencoder and the game training of generator and discriminator in AAE model.This is a semi-supervised classification method,which makes full use of data features and can effectively eliminate redundant information.By comparing the four experimental results,this algorithm can achieve better classification accuracy with fewer artificial label.2.A PolSAR image classification algorithm based on Auxiliary Classifier GAN(ACGAN)is proposed.In the preprocessing of PolSAR data,considering the spatial correlation between pixels,the neighborhood image block features are taken as the input data.Then,the ACGAN model is built.The parameters of the two networks are updated alternately during the training,which improves the ability of the discriminator to learn and capture the polarimetric characteristics.Finally,in the trained model,the auxiliary classifier in ACGAN is utilized as a classifier of sample set to obtain the classification labels of the whole image.By comparing the experimental results of four groups,this algorithm effectively improves the classification accuracy.3.A PolSAR image classification algorithm based on SLIC optimizing is proposed.Because AAE classification model is a classification algorithm based on a single pixel and ACGAN classification model is a classification algorithm based on a region block,this method applies different classification algorithms to different regions on the basis of SLIC superpixels segmentation.It improves the regional consistency of classification results and optimizes the boundary between different terrains.The rationality of this algorithm is verified by three comparative experiments.