Semi-Supervised PolSAR Terrain Classification Based On Mixup

Polarimetric Synthetic Aperture Radar(PolSAR)imaging is not affected by light,climate,etc.,and has a large number of applications in agriculture,military and civil fields.PolSAR terrain classification is an important application in the understanding and interpretation of SAR images.Due to the limitations of traditional methods and the advantages brought by deep learning,more research work has focused on the research of PolSAR terrain classification based on deep learning,which has led to the development of PolSAR terrain classification in a more accurate and faster direction.However,PolSAR terrain classification is a typical small sample problem.Because labeling PolSAR data requires a lot of field exploration,the cost of sample labeling is expensive,so that the number of labeled samples is very scarce,while the number of unlabeled samples is very large.Therefore,a large number of unlabeled samples could be used to assist in optimizing the deep learning classification model.Aiming at the small sample problem in PolSAR terrain classification task,this paper applies the mixup method to PolSAR terrain classification task,and studies the semi-supervised PolSAR terrain classification method based on mixup.The main research results are as follows:1.Terrain classification of small sample Po ISAR based on feature mixup.The original data augmentation strategy of mixup directly fuse images at the input of the neural network structure.Due to the existence of redundant pixels,a method of mixing up the features extracted by the shallow neural network is proposed and calculated consistency error loss using the KL divergence formula in order to learn the distribution of training data.In the semi-supervised classification task,in view of the low confidence of the single-branch prediction model,a multi-branch prediction structure model was designed.Using these methods and strategies has achieved very good results in the task of PolSAR terrain classification.2.Terrain classification of small sample Po ISAR based on regular terms.The loss function of the general semi-supervised classification method is usually composed of two parts:cross entropy and consistency loss(mean variance or divergence),but these methods do not constrain the distance between and within the classes.Based on this,a two-branch network model structure is designed,in which both branches use feature-mixup strategy.One branch is used to construct a triple or spatial neighbor relationship,and the other branch calculates the error loss based on the constructed triple or spatial relationship.The paper mainly applies the regular term of triples,the regular term based on pseudo labels,and the regular term based on embedded vectors to constrain the distance between and within classes.Compared with the more popular semi-supervised classification method in recent years,the use of the designed network structure and regular constraint items have improved to a certain extent.3.Terrain classification of small sample Po ISAR based on attention mechanism.In the neural network model structure,it is easy to lose some details due to the pooling(maxpooling or mean-pooling)operation.The attention mechanism modules are designed to extract useful details.The designed attention mechanism module can be directly inserted into the network structure based on feature-mixup.Firstly,a channel attention mechanism module is proposed.In the process of extracting features,different weight coefficients are set to different channels of the feature map;Secondly,different from the channel attention mechanism,in the process of extracting features,different spatial positions can also be set different weights.Based on this,a spatial attention mechanism module is proposed;Thirdly,a multi-branch prediction structure is combined to further improve the model's classification performance.