The Deep Self-paced Learning Approachs For Fully Polarimetric SAR Image Classification

Polarimetric synthetic aperture radar(PolSAR)has been one of the most important sensors in remote sensing.In addition to the day-night and all-weather advantages of SAR,PolSAR can transmit and receive electromagnetic energy in more than one polarization.This allows much richer characterization of the observed targets than single-polarization SAR.Therefore,PolSAR has been proven to be a valuable tool in many areas,such as military,agriculture and environment monitoring.PolSAR images classification is one of the most fundamental issues in the process of interpretation.In this paper,we commit ourselves to studying the two key issues of PolSAR image classification: classifier and feature extraction,which based on the self-paced learning(SPL)and deep learning.SPL has been empirically demonstrated to be instrumental in achieving a stronger generalization capability of classifier.Therefore,SPL can be used to improve the performance of classifier on the complex scenes of PolSAR images.And deep learning approach can converts raw PolSAR data into more abstract expressions through non-linear models.The main work of this paper are summarized as follows:(1)To obtain better classification results on the complex scenes of PolSAR images,a novel support vector machine algorithm based on self-paced learning(SPSVM)for PolSAR classification is proposed.Support Vector Machine(SVM)is used to train classifier under the guidance of the SPL which learns the easier samples first and then gradually involve more difficult samples in the training process.This learning mechanism can improve the generalization capability of classifier.Moreover,in order to improve the performances of classifier further,the neighborhood information is introduced to improve the calculation formula of samples' weights.Experimental results on three real PolSAR images show that the proposed method performs well especially on some complex scenes.(2)Using traditional methods to extract features such as polarization features or texture features of PolSAR data has limitations.The polarization features or texture features cannot represent the nature of the raw data.Aiming to solve this problem,a self-paced stack autoencoder(SPSAE)for PolSAR images classification is proposed.The stack autoencoder network(SAE)can not only extract the more abstract features from the raw PolSAR data,but also has the advantage of easy training.In addition,the learning mechanism of self-paced learning can help the network to avoid the local optimal solution and achieve to a better generalization result.The experimental results show that the proposed method has the advantages of easy training and strong generalization capability even when the number of labeled samples is relatively small.(3)In order to fully exploit the information between the four channels(HH,HV,VH,VV)and the spatial information of PolSAR images,a PolSAR images classification method based on self-paced convolutional neural network(SPCNN)is proposed.In our method,we construct a 6-layer convolutional neural network(CNN),and each pixel is denoted by a three-dimensional tensor block formed by its scattering intensity values on four channels,Pauli's RGB image and its neighborhood information.Then the CNN is used to extract the abstract channels-spatial features for classification.When train the convolutional neural network,the self-paced learning method is used to help the network converge to a better solution and improve the generalization of the network.The experimental results show that the proposed method can obtain the more smooth classification results.