PolSAR Image Classification Based On Deep Complex-Valued Neural Network

Polarimetric synthetic aperture radar(PolS AR)emits and receives multiple polarization electromagnetic waves during the imaging process,and the pixel is no longer backward scattering coefficient,but a complex-valued polarization scattering matrix or coherency matrix that can reflect the polarization scattering characteristics of the targets,thus enabling better acquisition of target information.With the application of PolSAR systems,massive PolSAR data have been widely used in both military and civilian applications.The PolS AR classification is a key problem in PolSAR image interpretation,aiming to assign targets to a certain category based on the scattering characteristics.Deep learning algorithms adaptively extract features from PolSAR images and have stronger comprehensive classification performance than traditional methods.Most methods are based on conventional real valued deep neural networks,which results in a weak ability in processing complex-valued information.In this dissertation,we study the complex-valued neural networks for PolSAR image classification,and develop the PolSAR image classification algorithm in terms of improving the spatial and scattering information capture capability,improving the intra-class and inter-class metric,and improving the randomness of the network.This dissertation consists of following five aspects.In the first part,the basic complex-valued data structure of PolSAR and the basic component of complex deep learning are expounded,which is the basis of the whole dissertation.On this basis,a non-local block for complex-valued networks is designed by embedding the non-local method to address the insufficient spatial association in the Convolutional Neural Networks(CNN)for PolSAR image classification.The output of this module is a weighted sum of the pixel and other pixel within the feature map.The block establishes the spatial dependencies among all pixels within the feature maps.Based on this block,a complex-valued non-local neural network(CNLNN)is proposed for PolSAR classification.The feasibility and effectiveness of this approach are experimentally verified.In the second part,a deep complex-valued 3D neural network(CV-3D-CNN)is proposed to enhance the CNN for multi-polarization channel feature extraction capability and used for PolSAR classification.Most CNN-based models are currently limited to processing 2D real values inputs,and thus cannot effectively extract the physical scattering mechanisms contained in the complex-valued covariance/coherence matrix in PolS AR data.To this end,CNN structure is improved by designing a complex-valued 3D convolution,a complex-valued 3D pooling layer,and a fully connected layer.Compared with CNNs,CV-3D-CNN captures deep semantic information from adjacent resolution cells by performing 3D complex-valued convolution in both spatial and scattering dimensions simultaneously.The experimental results validate the effectiveness and better classification accuracy of the CV-3D-CNN.In the third part,the HCRF-CV3DCNN model is proposed to combine the learning capability of CV-3D-CNN for semantic features and the modeling capability of random fields for spatial statistical features of PolSAR images.This part proposes a hybrid conditional random field based on a CV-3D-CNN for PolSAR image classification,named HCRF-CV3DCNN.HCRF-CV3DCNN extracts the deep features of PolSAR images by CV-3D-CNN,thus utilizing the magnitude and phase information of PolSAR data to produce the class probability for the random field framework.Furthermore,based on the class probabilities from CV-3DCNN,the relative entropy of the class distribution is derived to adjust the label interactions to improve the accuracy of edge localization in classification.Finally,to capture PolSAR image information more completely,the depth features and PolSAR scattering statistics are integrated into the Bayesian fusion-based HCRF-CV3DCNN.In this way,HCRF-CV3DCNN effectively combines the representation learning capability of deep learning models with the modeling capability of random fields including spatial correlation and data statistics.The experimental results demonstrate that HCRF-CV3DCNN is an effective PolSAR image classification method combining conditional random field and deep learning models,which has superior performance over the recent deep learning models.The fourth part aims to strengthen CNN for inter-and intra-class similarity measures and increase the classification performance under small sample conditions,thus designing a weight-sharing metric network for PolSAR feature extraction and classification.CNNs generally focus on the correlation between pixels and labels but have fewer constraints on interor intra-class feature distances.First,a triplet CV network(TCVN)is proposed to learn CV representations from PolSAR data by maximizing the inter-class distance and minimizing the intra-class distance.It uses CV convolution and CV Euclidean to maintain the phase components,and applies CV-dropout and CV L2 parameter regularization to reduce overfitting and further improve the performance.Subsequently,CV K Nearest Neighbor(CV-KNN)calculates the distance of CV representation and distinguishes similar pixels.CV-KNN couples well with TCVN because they are both based on the Euclidean distance in the complex domain.TCVN can extract hierarchical features by comparing PolSAR data in the complex domain and maintain the phase component by performing CV calculation.Experiments on PolSAR images demonstrate the effectiveness of CV Euclidean distance and show that TCVN can process PolSAR data more efficiently and achieve comparable performance in PolSAR image classification even with smaller datasets.In the fifth part,to improve the ability of CNN against overfitting and to enhance the discrimination of uncertainty,randomness is introduced to CNNs,and a Complex-Valued Variational Inference Network(CVIN)is proposed for PolSAR image classification.The weights of neuron in CNN s are constant,which may lack randomness.Different from CNN,the complexvalued neurons in CVIN are no longer fixed values,but Gaussian distribution in the complex domain.In this way,CVIN becomes a flexible and uncertain network structure,which can alleviate the overfitting phenomenon caused by the complex imaging mechanisms and the random scattering noises in PolSAR images.CVIN uses a novel evidence lower bound(ELBO),which enables variational approximate inference of the network weights distribution through backpropagation.CVIN can generate approximation posterior distributions of labels given the input data and make label predictions.Experiments on real PolSAR images validate the feasibility of CVIN and illustrate the potential of CVIN as a competitive method for PolSAR classification.