Unsupervised Feature Learning And Fusion Classification Methods For SAR/PolSAR Images

Synthetic aperture radar(SAR)obtains high-resolution images of target scenes through electromagnetic wave transceivers and imaging processing.Based on traditional SAR,polarimetric synthetic aperture radar(PolSAR)alternately changes the polarization patterns of signal transceiving and transmitting,which can obtain richer target scattering information.SAR/PolSAR images contain rich and complex information,which can be revealed by image segmentation and classification to achieve automatic interpretation of high-resolution terrain information.SAR/PolSAR image segmentation with high accuracy can accurately invert the target types and spatial distribution in the observed scenes,providing key information for military terrain detection,natural disaster assessment,crop yield estimation,etc.Recently,classification methods based on deep learning have attracted much attention in SAR/PolSAR interpretation.Deep learning-based SAR/PolSAR image classification tasks generally require prior label information.A suitable network is then trained to automatically predict the label of each pixel on the basis of input measurement data.However,labeling is time-consuming and needs expert knowledge.In addition,due to the unique imaging mechanism,SAR/PolSAR images contain inherent speckle noise,resulting in a wide range of pixel values in homogeneous regions and edge distortion in heterogeneous regions.How to design the network structure to extract effective discriminative deep features using a limited number of labeled samples is of great significance for practical applications.Based on the national natural science foundation of China,this project investigates unsupervised feature learning and fusion classification methods for SAR/PolSAR images.The focus of this thesis is on SAR images with complex scenes and complex-valued PolSAR data.Based on convolutional neural networks(CNN),enforcing population and lifetime sparsity(EPLS),and complex-valued CNN,generalized networks trained with a small number of labeled samples are designed to extract discriminative features.The specific innovative work is as follows:1.To overcome the shortcomings of traditional unsupervised deep models in capturing fine structures,the 3-dimensional sparse model(3-DSM)is proposed to improve the performance of unsupervised discriminative feature extraction and classification.3-DSM defines the sparsity of convolution kernels,in which each convolution kernel is unique to capture specific structural information of SAR images.As a result,the redundancy of convolution kernels is reduced.In addition,the 3-D sparse maps are constructed based on the EPLS model.The sparse convolution kernel is learned by minimizing the error between the feature maps and the 3-D sparse maps.Then,discriminative features are extracted in an unsupervised way.The experiment results on the simulated and real SAR image demonstrate that 3-DSM can learn sparse convolution kernels in an unsupervised way and then extract discriminative features.Besides,3-DSM achieves precise segmentation of complex structures and improves the accuracy of classification.2.Existing unsupervised models extract the deep features of PolSAR images by reconstructing the inputs,which neglects the discriminant of the features that facilitate classification.To address this problem,the complex-valued enforcing population and lifetime sparsity(CVEPLS)model is proposed.Firstly,CV-EPLS defines an activation metric to measure the difference between complex vectors to activate the complex sparse matrix and ensure its strong population sparsity,.In addition,CV-EPLS designs a grid search strategy to ensure the even distribution of activations in the complex sparse matrix,thus improving the differentiation between sparse features.Finally,complex sparse matrices fulfilling population and lifetime sparsity are used for unsupervised optimization of the complex convolution network,which fully integrates the amplitude and phase information of different polarimetric channels.Besides,CV-EPLS extracts discriminative deep complex features without labeled samples and avoids overfitting.Experimental results on the classical PolSAR dataset confirm that CV-EPLS can extract discriminative deep complex features in an unsupervised way.These features promote the preservation of fine structural information and smooth homogeneous regions,thus avoiding overfitting and improving classification accuracy.3.In general,it is difficult for a network with single receptive fields to segment SAR images of complex scenes well.Besides,the underutilization of labeled samples is another difficulty.To overcome these problems,the hierarchical fusion CNN(HIFCNN)model is proposed,which extends the network width to extract hierarchical features and capture more comprehensive information.HIFCNN designs three sub-networks with different receptive fields to extract hierarchical features,thus can capture both fine structure and regional information of SAR images.Then,Dempster-Shafer evidential theory(DSET)is applied to integrate these hierarchical features.In addition,the three sub-networks are trained in parallel using the same sample set,and the limited samples are fully utilized.The experiment results of the simulated and real SAR images with complex scenes demonstrate that HIFCNN can capture fine structures while maintaining smooth homogeneous regions and is more robust against coherent speckle noise.4.For pixel-based deep models,the spatial relationship of PolSAR images is often ignored.In addition,the underutilization of samples is another problem.In order to overcome the above problems,the semi-supervised complex network with spatial statistics fusion(SCNSSF)is proposed,which improves classification performance using unlabeled samples.First,the SCN-SSF updates the pseudo labels of the unlabeled samples iteratively during training,whose errors constitute the regularization term of the objective function.Thus,the model alleviates the overfitting of the network with a small number of labeled samples and improves its generalization.In addition,SCN-SSF models high-order neighborhood label interactions based on product-of-experts(POE)to obtain contextual information,making full use of spatial relationships to enhance label consistency and improve the ability against speckle noise.Further,DSET is used to achieve effective integration of pixel-level and contextual label information,which can preserve structural infromation and improve the smoothness of classification.Experiment results on the classical PolSAR dataset show that the SCN part of the SCN-SSF can alleviate overfitting and improve generalization using PolSAR unlabeled samples.The subsequent SSF part can effectively correct misclassification to improve the smoothness of homogeneous regions and the ability against speckle noise.