Research On PolSAR Image Classification Based On Fully Convolutional Network And Graph Embedding

Nowadays,with the rapid development of artificial intelligence,how to apply the advantages of deep learning and big data to PolSAR images classification is a hot topic in the field of remote sensing.Feature extraction or learning is a key step in PolSAR image classification.One of the current research trends is the effective extraction(or learning)and fusion of polarimetric features and spatial features,where the target decomposition based extraction of polarimetric feature has become mature.While how to optimally represent the extracted high-dimensional polarimetric features,how to adaptively learn the discriminative spatial features,and how to effectively integrate polarimetric information with spatial information to maximize the accuracy of classification are the core issues at the moment.Inspired by enormous success of fully convolutional network(FCN)in semantic segmentation,as well as the similarity between semantic segmentation and pixel-bypixel polarimetric synthetic aperture radar(PolSAR)image classification,exploring how to effectively combine the unique polarimetric properties with FCN is a promising attempt at PolSAR image classification.Moreover,recent research shows that sparse representation,low-rank representation,and manifold learning can convey valuable information for classification purposes.Considering that there are few correctly labeled PolSAR data,which makes it difficult to train a good deep network for pixel-by-pixel PolSAR image classification,this paper proposes a PolSAR image classification scheme combining shallow learning and deep learning,in which the shallow subspace features constrained by sparsity,low-rankness and manifold are integrated with the deep spatial patterns automatically learned by FCN for classification.The main contributions of this paper are summarized as follows:(1)a shallow subspace learning based on graph embedding is introduced to capture the essential structures of high-dimensional polarimetric data.Under the framework of graph embedding,sparse representation,low-rank representation,and manifold can be merged to retain the important properties of low-dimensional representations,such as local features,global features,and local intrinsic geometry and so on;(2)a pre-trained FCN-8s model on the optical image dataset is transferred to extract the nonlinear deep multi-scale spatial information of PolSAR image;(3)the shallow sparse,low-rank,and manifold subspace features are integrated with deep spatial features by a weighted way,thus can boost the discrimination of fusion features for subsequent classification,because of their complementary advantages;and(4)in order to more effectively integrate polarimetric information with spatial information,this paper further use multiple parallel FCNs(MFCN)to learn deep spatial features,and achieve adaptive fusion of spatial-polarimetric information simultaneously.In order to remove redundant information,the higher-dimensional spatial-polarimetric feature,stacked by the feature maps outputted from MFCN,are fed into a manifold graph embedding to obtain the features with stronger ability of representation and discrimination in the manifold subspace.And then,the integrated features are combined with a discriminant model to obtain the final classification result.Finally,extensive experiments on three real PolSAR datasets indicate that the proposed method can achieve competitive performance.Particularly in the case where the available training samples are limited,the synergy between shallow learning and deep learning can make their advantages beneficial for each other to maximize the classification accuracy.