PolSAR Image Classification Based On Manifold Network

Synthetic aperture radar(SAR)can provide abundant information for a vast ma-jority of application areas due to its ability to work day and night under all weather conditions.Polarimetric SAR(PolSAR)can further reveal the differences among differ-ent land scatters thus enhancing the application ability of radar.Existed PolSAR image classification methods are mostly based on feature engineering with fixed models,which has the limitation of poor data adaptability and low information utilization rate.A promising solving direction is to combine SAR interpretation with deep nets capable of feature learning.However,when applying deep learning networks designed for optimal images on SAR images,problems like over fitting,poor generalizing may easily occur.How to concur these problems becomes an unavoidable difficulty in feature learning based SAR classification.Considering that the feature learning process of deep networks can be viewed fruit-fully as a process of approximating a mapping function which maps the original data to its corresponding label,manifold hypothesis is introduced in our work to improve the efficiency and effectiveness of the network mapping.Manifold hypothesis claims that real world data presented in high dimensional space is likely to concentrate in the vicinity of non-linear sub-manifolds of much lower dimensionality.Points of different classes are likely to concentrate along different sub-manifolds,separated by low density regions generated by noise in data.In this framework,the classification utilizing man-ifold learning manages to better model the core information of original data so as to separate different sub-manifolds corresponding to different classes.Compared with con-volutional mapping in common networks,this kind of manifold mapping can construct a more efficient classification mapping thus better realizing the cognition of targets and in the meanwhile simplifying the structure of network.The main contribution of this work is threefold.1)Based on the nature of clas-sifcation,the details about how convolutional module and manifold module work in a classification process are discussed,where a visualization experiment is also provided to verify the actual mapping results.2)With the motivation of improving the network mapping by manifold learning,a manifold network for PolSAR classification named manifold embedding network is proposed.This net embeds manifold module in fully convolutional network to deal with the high-dimensional low-level feature by manifold mapping,which automatically maps the original data to a low-dimensional space com-posed of the core dimensions.Data in this core feature space is consequently input into the following classification net to continue the classification mapping.3)Another mani-fold net named manifold mapping network is further proposed after manifold embedding network,which takes advantage of label information to improve manifold embedding mapping and thus improve the following network mapping.Finally,the proposed networks are validated in three real polarimetric SAR data sets and a series of contrasting experiments were conducted,including the compari-son between low-dimensional feature and high-dimensional feature,and the contrast between the common net and the proposed nets.The experiment results verify the effectiveness and feasibility of the proposed networks in this dissertation.