| Polarimetric Synthetic Aperture Radar(PolSAR)imaging is one of the most advanced means of earth observation.It has been widely used in military,agriculture and other fields because of the characteristics of obtaining observation information without the limitation of weather,day and night.Terrain classification is an important part of PolSAR image interpretation.With the development of deep learning,many scholars have paid attention to PolSAR terrain classification based on deep learning in recent years.Deep learning relies on large-scale labeled data,and the label of PolSAR data usually requires a lot of manpower and material resources,which affects the progress of PolSAR related tasks.In recent years,with the development of remote sensing technology,more and more unlabeled PolSAR data are available.It is becoming more and more important to use a large number of unlabeled samples and a small number of labeled samples to train a high-precision PolSAR terrain classification model.This thesis studies the PolSAR terrain classification based on deep semi-supervised learning.The main research results are as follows:1.Semi-supervised learning methods based on pseudo-label mainly focus on the relationship between labeled samples and unlabeled samples;semi-supervised learning methods based on consistency regularization mainly focus on the relationship between the same unlabeled samples after different perturbations.Both of these two types of methods ignore the relationship between unlabeled samples.This thesis designs two regularization loss functions to constrain the model by exploiting the relationship between unlabeled samples.One encourages that the relationship between low-dimensional features of unlabeled data is similar to the relationship between their class probability features;the other encourages the relationship between unlabeled sample low-dimensional features to be consistent before and after perturbation.In addition,a pseudo-label mechanism based on the feature memory bank is designed,which uses the similarity between the low-dimensional features of the current data and the features in the memory bank to correct the category probability of the current data to improve the accuracy of the pseudo-label.2.Traditional Co-Training methods require data to have two independent views.However,such data is rarely available.Some scholars propose to use two neural networks with different initialization methods to perform Co-Training on data with a single view.However,the difference between the two models with different initialization methods is too small,and they will gradually converge during the training process,thus losing the meaning of CoTraining and mutual learning.Based on the traditional Co-Training method,this thesis proposes Co-Training based on different classification strategies.Two models are trained with different training strategies.Among them,model A is a traditional classification network structure;while model B only contains a feature extraction network,which is trained by metric learning.In addition,in order to alleviate the impact of pseudo-label bias,the classification of labeled samples and unlabeled samples is deconstructed,and an additional classifier is added for unlabeled samples.Finally,a consistency loss is designed to make the two networks interact during training.3.In the real situation,the labeled samples obtained from PolSAR data and the samples to be classified cannot satisfy the independent and identical distribution assumption,which makes most supervised and semi-supervised methods unable to achieve the expected results.In response to this problem,this paper designs a semi-supervised domain adaptation method based on adversarial,which uses the principle of generative adversarial to align the features of labeled samples and unlabeled samples,which alleviates the decline of the accuracy of the model.In addition,for the problem that labeled samples are far less than unlabeled samples,mixup is used to mix labeled samples and unlabeled samples to generate data in multiple intermediate domains,reducing the difficulty of domain alignment.Finally,a pseudo-label selection strategy based on gaussian mixture model is designed to accurately select unlabeled samples with correct pseudo-labels for unsupervised loss calculation. |
