| Polarimetric SAR images(Polarimetric SAR,Pol SAR)are high-resolution remote sensing images formed by the backscattering of polarization waves from objects.They are widely used in geological exploration,vegetation growth assessment,urban planning,and ocean monitoring.Pol SAR image classification is one of the important links of polarization information processing and object interpretation.Recently,thanks to the development of deep learning technology,many new ideas and new models have emerged to solve the problem of Pol SAR image classification.The encode-decode network structure of the fully convolutional network is very suitable for the classification of Pol SAR images,due to the inherent defects of the fully convolutional network and the particularity of the problem itself,there are still some problems to be solved.It is difficult to obtain large amounts of labeled data of Pol SAR images,and the semi-supervised classification method has become one of the hotspots of the current research.Based on fully convolutional networks and semisupervised learning,this paper proposes several high-accuracy,high-efficiency Pol SAR image classification methods.The specific content is as follows:(1)Aiming at the problem of efficient encoding and representation of high-dimensional features of Pol SAR images and the classification ambiguity caused by loss of detailed features in up-sampling and decoding,a Pol SAR image classification method based on the feature attention and feature refinement fully convolutional network is proposed in chapter 2.This method introduces two attention mechanism models and a residual refinement model in the network.The dual feature attention model adaptively obtains more robust feature encoding of the high-dimensional feature of Pol SAR images.And the fusion feature attention model obtains the consistency of fusion features at different stages.The residual refinement model is combined to reconstruct consistent detailed features.This method enhances functions of feature representation and feature reconstruction of the fully convolutional network.Experiments verify that it can significantly improve the classification accuracy of Pol SAR images.(2)Aiming at the problem that the edge region of each category is not finely classified and the region consistency of the classification is not strong,a Pol SAR image classification method based on the edge-aware dual branch fully convolutional network is proposed in chapter 3.This method adds an edge-aware network to the structure of the fully convolutional network in chapter 2 so that forms a dual branch fully convolutional network.With the synergetic training of two network branches,the purpose of constraining edges of the classification region and enhancing the consistency of the classification region are achieved.Edges of categories are subtly classified by this method,which enhances the region consistency of the classification.Experiments verify that it can effectively improve the classification accuracy of Pol SAR images.(3)Aiming at the small sample problem in Pol SAR image classification,a semi-supervised Pol SAR image classification method based on adversarial training is proposed in chapter 4.This method adds a discriminative network to the structure of the fully convolutional network in chapter 2.The two networks utilize adversarial training to gradually correct the high-order inconsistency between the classification result and the true label,so that the prediction result is closer to the true value.In addition,the method reasonably defines the classification loss of unlabeled samples,and the prediction confidence of the discriminant network is used to automatically label unlabeled samples and dynamically expand the training set,so that the information of a large number of unlabeled samples can be effectively used.Experiments verify that it can improve the classification accuracy of Pol SAR images to a certain extent by using only a small number of labeled samples. |
