Semi Supervised Polarimetric SAR Image Classification Based On Self-Correcting Training And Sample Selection Strategy

PolSAR classification plays an important role in military observation,urban planning,crop classification and various civil fields.However,due to the high cost of supervised classification,the effect of unsupervised classification is not ideal.Therefore,there are more and more researches on PolSAR images classification using semi-supervised method.A small amount of labeled data and a large number of unlabeled data are combined to improve the classification results.But the problem of semi-supervision is to ensure whether the selected samples are reliable,how to avoid the impact of wrong samples on the model,and how to judge whether the model performance is good or bad.Therefore,this paper starts with the strategy of selecting samples and the training strategy of unlabeled samples.Firstly,the basic theory of PolSAR image,the basic situation of deep learning,the main methods and research status of semisupervised learning are introduced.Then,the main research contents are as follows:1.For the training strategy.A PolSAR images classification method based on SCDS-Net(Self-Correcting Depthwise Separable Network)and selfcorrection strategy is proposed.The training mode uses the self-correction strategy.At the same time,the center loss function is introduced from the point of view of sample selection and Label Smooth loss function is introduced from the point of view of unlabeled sample training.Finally,the network model is constructed on the basis of depthwise separable convolution,and an effective feature extraction model SCDS-Net is designed by asymmetric convolution.The experimental results show that the method can achieve a good classification effect in the case of limited number of real labeled samples.2.For how to select reliability samples.A PolSAR classification method based on affinity matrix and loss function is proposed.The training strategy is self-training.At the same time,a loss function suitable for unlabeled sample training is designed according to affinity matrix.Finally,in order to enhance the similarity of adjacent pixels,conditional random field is used to optimize the prediction results.Three groups of polarization data are used in the experiment,and the method in the chapter can achieve better classification results in the case of a limited number of real labeled samples.