PolSAR Classification Based On Semi-supervised Generative Adversarial Networks

PolSAR is currently an imaging radar system which is used extensively.PolSAR data contains masses of polarization information on various terrain objects,thus suitable for postprocessing of remote sensing data.Therefore,its related classification work has always been a hot research topic in the field of radar imagery.Deep learning algorithm can learn features from the input data autonomously.Compared with the traditional artificial feature extraction methods,it can save a lot of labor cost and time,and get better classification results by using a small-sample training classifier.In recent years,deep learning technology has drawn wide attention in the field of remote sensing and has achieved good results.Among those deep learning methods,generative adversarial networks(GAN)is quite suitable for the terrain classification work due to its advantages of generating data to assist the training classifier and being more robust to noise.This paper mainly studies the semi-supervised PolSAR terrain classification method based on GAN.The main research work is as follows:First,a semi-supervised GAN is proposed to address the problems regarding the small sample in the field of PolSAR.This method integrates the idea of semi-supervised learning to improve the original GAN.Only a small number of labeled samples are needed for the network to learn the abstract features of the data,and a sufficient number of unlabeled samples are also used to optimize the learning ability of the network.The PolSAR data processed by this model are a series of image patches.The finished model can produce a good classification effect.Its good classification performance can be proved by the experiments.Second,an improved end-to-end semi-supervised GAN is proposed to solve the problems such as complex preprocessing involving multiple steps and slow model training.The improved model integrates the fully convolutional algorithm into the semi-supervised GAN,thus implementing the pixel-wise end-to-end PolSAR terrain classification tasks.In this algorithm,data preprocessing is simplified,the space required for model training and sample testing is minimized,and the amount of calculation is reduced,so the training and testing can be performed more effectively.The experiments have proved that the classification performance of the model is improved,the training time is significantly shortened,and the test efficiency is higher.Third,a conditional semi-supervised GAN is proposed to solve the problem of generating samples with unrestricted conditions.The proposed algorithm adds labels to the network as a constraint to improve the end-to-end semi-supervised GAN.The generated samples in the original model don't have any label,and their effects are limited.Therefore,we input the label information into the network to guide the training process and generate data that are closer to the realistic distribution.At the same time,the discriminator is provided with more data for the purpose of training and achieving better classification results.Experiments show that the classification performance of the model has been improved.