Research On PolSAR Image Classification Based On Fully Convolutional Networks And Manifold Learning

Polarimetric Synthetic Aperture Radar(PolSAR)system can transmit and receive a variety of electromagnetic waves at the same time to survey the polarization scattering characteristics of ground objects.The scattering matrix measured by PolSAR contains abundant target information,so it is widely used in the field of terrain classification.According to the existing research,there are some problems to be solved in PolSAR image classification.On the one hand,it is difficult for traditional manual features adapt to the data automatically,fixed pattern composed of feature extraction and classifier can hardly achieve good results.On the other hand,deep learning method designed for optical images is prone to over-fitting and has poor generalization ability when faced with PolSAR data.Therefore,how to design an efficient feature description and achieve accurate terrain classification with small samples is a very challenging problem in current PolSAR image classification field.For polarimetric terrain classification,feature learning based on deep network can realize underlying feature abstraction through automatic learning,thus to complete the mapping of original data to label.However,this kind of method only uses partial polarimetric information,and it is difficult to describe all ground types completely.Therefore,combining deep network with polarimetric features becomes a new breakthrough direction.High-dimensional polarimetric features extracted by multiple target-decomposition methods can fully represent the PolSAR image,but there is some redundancy.Manifold learning can model the core variables of high-dimensional data and mine its eigenstructure as much as possible,to achieve the purpose of separating different data types.Therefore,this paper,takes manifold mapping as the starting point,deeply explores PolSAR image classification methods based on manifold learning and Fully Convolutional Networks(FCN).In this paper,the main contributions are as follows:(1)in order to obtain feature with strong adaptability and high utilization,this paper adopts the strategy of transfer learning,and combines it with FCN model to automatically learn the deep multi-scale spatial information for PolSAR imagery,breaking through the limitation of manual features.(2)To enhance feature separability,nonlinear manifold learning is employed to map high-dimensional polarimetric features to eigenspace which can represent the original data better.Meanwhile,the shallow manifold representation is weightly embedded into the deep multi-scale spatial features of FCN model,makeing each other complementary and boosting fusion feature's discrimination ability.(3)Considering that the non-explicit manifold learning method has “external samples” problem and can not be back-propagated,this paper further introduces the grassmann manifold network,and integrates FCN's deep features to form an integrated manifold network.Finally,a series of comparative experiments are carried out on three real PolSAR datasets,including the contrast fusion of polarimetric coherence feature with deep spatial information,grassmann manifold network with feature fusion in data layer and classification layer and etc.The experimental results have demonstrated that the algorithm proposed in this paper is effective and superior for PolSAR imagery classification.