POLSAR Image Classification Via DKSVD And AE Network

Polarimetric synthetic aperture radar(POLSAR)image terrain classification is a research topic of interest being explored for many years.Through three decades of researching on POLSAR image classification,many methods have been proposed by different researchers.With the development of sparse representation and neural network,numerous classifica-tion methods of machine learning fields have been successfully applied on POLSAR im-age.Compared with the natural pictures,POLSAR image has many special characteristics of itself.For example,the statistical distribution of POLSAR images is different from natural pictures.Consequently,the POLSAR classification results obtained by many mature meth-ods(proposed in the natural image processing community)are not satisfactory.Recently,researchers introduce an ocean POLSAR image classification approaches to overcome dif-ferent technical fortresses,ranging from the traditional classification methods to the machine learning.Based on the background mentioned above,this thesis makes a thorough study on POLSAR image classification task,and has achieved some success,which are summarized as follows:1.Based on the background that the sparse representation classifier(SRC)has not been widely used in POLSAR image classification,the paper proposed a POLSAR image clas-sification method based on the combination of discriminative dictionary learning(DKSVD)and non-subsampled contourlet transform(NSCT).When extracted the classification fea-tures of POLSAR image,we use NSCT to transfer the constructed feature images on origi-nal data domain and obtain the feature images on transform domain.Then,we abandon the high-frequency feature images and only extract the low-frequency feature images of trans-form domain as the image classification features.The low-frequenciy coefficients of NSCT domain contain the main useful information,and the high-frequency coefficients of NSC-T domain include lots of noise of image.Thus,the low-frequency coefficients have better discrimination ability than other high-frequency coefficients,and are beneficial to image classification.Using the extracted classification features to initialize the DKSVD and com-bine the labels of samples to train this model and then using the trained DKSVD model to achieve the classification task of whole POLSAR image.2.Considering the statistical distribution of POLSAR image,we propose an effective image classification method.Combining the Wishart distance with Auto-encoder(AE)network model to form a new type of auto-encoder network which named Wishart-AE(WAE)net-work model.The network model improve the separability of classification features of POL-SAR image because the new model considers the statistical distribution characteristics of POLSAR image which using Wishart distance to measure the similarity of POLSAR image pixel.Compared with AE model,the WAE model is contribute to the classification perfor-mance of POLSAR image.The effectiveness of this method has been verified on different POLSAR images.3.To make good use of spatial information,the paper proposed the Wishart-convolutional auto-encoder(WCAE)model which combing the neighborhood information to improve the performance of image classification.The basics of WCAE model is convolutional auto-encoder(WCAE)network,which applying the two-dimensions convolution through the process of image feature extraction that could take full advantage of information of neigh-bourhood pixel.Based on the CAE network,the paper considered the special statistical distribution of POLSAR image,and combined the Wishart distance measurement into the training process of the CAE network.Compared with the CAE network,WCAE model can achieve higher classification accuracy because this new network could obtain the image classification features,which are more suitable for POLSAR data.Because the procedure of feature extraction of WCAE model is unsupervised,which only need a small number of class label during the training process could obtain satisfactory result,the training pro-cess is much time-saving.Given the above,our method not only improves the classification performance but also saves the training time.4.Considering the clustering algorithms could explore the label information automatical-ly,we proposed a classifier model with label discriminating ability.We add a traditional clustering algorithm into the objective function of WAE model and propose the Cluster-ing Wishart-auto-encoder(Clustering-WAE)model.The Clustering-WAE model accom-plish the clustering process of hidden layer data during optimizing the objective function of WAE model.By minimizing the distance between the sample and its cluster center,the network weight could be optimized.The Clustering-WAE model have the distinguishing performance when extracting the classification features of POLSAR image.Finally,we con-nect the Clustering-WAE with Softmax classifier to achieve the image classification task.The result show that the proposed method improve the classification performance of whole POLSAR image.