Polarimetric SAR Images Classification Based On Discriminative Feature Learning

Polarimetric synthetic aperture radar(PolSAR),which is used to detect targets on the ground or in space,has the characteristics of all-weather,all-day,high precision and high efficiency.Due to its unique ability of transmitting and accepting multiple combinations of polarized electromagnetic waves,rich polarization information involved in the detected targets,which is conducive to the subsequent data analysis.Image understanding and interpretation includes lots of tasks,such as target recognition,transforming detection,noise reduction and so on.Classification of the Pol SAR image is one of the most important applications in Pol SAR image understanding and interpretation,and it is also the basis of object identification.Classification result can be output to the users as the final result to provide important target information,and it also can be used as intermediate result to provide useful image information for some subsequent work,such as target recognition,edge detection and so on.However,disadvantages exist in polarimetric SAR images,such as difficulty of obtaining discriminative features for classification,large data scale,lack of training samples with labels,and serious speckle noise,which makes the classification uneasy.The dissertation is a research on Pol SAR data,in view of the probelms in Pol SAR image classification,including the difficulties in obtaining suitable features,the large amount of samples with noise and the small amount of samples with lables,a series of novel,robust and intelligent classifiers combined with discriminative features are designed.The contributions are as follows.1.Considering that it is difficult to obtain smooth regions and clear edges simultaneously in Pol SAR image classification,a novel classification scheme which combines a high probability selection and an adaptive MRF is designe.SVM classifier is a discriminative traditional classifier,but this kind of classifier does not consider the spatial information of Pol SAR images.To solve this problem,the Wishart distance is applied to enhance the discriminative of probabilistic output SVM classifier,and at the same time,an adaptive window based MRF is introduced for smoothing the homogeneous regions.Experimental results indicate that the classifier can get clearer edges and smoother homogeneous regions at the same time,which improves the classification accuracy.2.Aiming at the lack of training samples of Pol SAR images,a Pol SAR image classification method that combines sparse autoencoder(SAE)and edge preserving Wishart Markov random field is proposed.SAE is an end-to-end learning framework of feature classifier,the learning process is guided by the classification.Based on the SAE,discriminative features which are beneficial to the classification can be learned automatically without supervision,and very few training samples are necessary for fine-tuning,it makes sence in reducing the need for labeled samples to a large extent.Then,an edge preserving MRF is used for reducing the speckle noise and compensating the uncertainty and ambiguity of the SAE classification results.Meanwhile,an error correction strategy is proposed to correct the pixels that are misclassified by the previous MRF.The classification framework consists of two layers.The first layer leads the initial classification result based on the SAE and offers abundant information for improving the classification of the boundary regions.The second layer is designed on the basis of the spatial information of Pol SAR,for revising the classification result of the first layer by using MRF.The validity of the proposed method has been verified by different Pol SAR data.3.Taking the problem that the classification reults of two-step classifiers heavily dependent on the classification reult of the first step,a discriminant classification model is proposed.The characteristic function which is suitable for Pol SAR image classification is redesigned under the Max-Margin framework.The characteristic function consists of two parts,scattering term and spatial term.The scattering term is designed by the classical discriminative SVM algorithm,which can learn the decision boundaries under the Max-Margin framework effectively.Besides,the accuracy of Pol SAR images classification can be improved in the case of limited training samples.The conditional random field is applied in the spatial term to fuse the context information into both the observation domain and the label domain,and the Wishart distribution is used to describe the statistical characteristics of the Pol SAR image.In this method,the spatial information is utilizied in the learning of the classifier,unlike the previous two-stage classifier,which makes use of the spatial information for revising the misclassification.Experiments show that the proposed framework can achieve high classification accuracy and clear classification images.4.Considering the limited of the training samples and the effect of speckle noise in Pol SAR images,which further affects the learning performance of the classifier,a recursive convolution neural network model(CNN)is preserved.Samples with high confidence of each classification result will be used as the training samples of the next training.Then,a semi-supervised model is obtained for Pol SAR image classification.This model is independent of the dependence of supervised classification on manual calibration samples.Furthermore,the model is an end-to-end classification framework based on discriminative feature learning which can learn the spatial texture features of polarized SAR images automatically while performing convolution operations in CNN.In addition,this model tries to learn features that are beneficial to classification from highly confident samples.There are three advantages of the proposed model: firstly,the problem of small samples is solved by increasing the training samples from each iterative classification result.Secondly,the low confidence samples are removed in each iteration to reduce the impact of noise samples on the robustness of the model.Finally,the initialization of CNN parameters in each iteration process is based on the results of the previous learning.As a result,the parameters will be set more and more robustly,so that the entire model will not have poor performance due to random initialization.