Statistical And Robust Multilayer Network For High-resolution SAR Images Classification

Synthetic Aperture Radar(SAR)is an active remote sensor for Earth observation,and its operating frequency bands are in the range of microwaves.SAR can acquire rich information about the Earth surface with the coherent imaging mechanism,and it has been widely used in the fields of military and civilian.SAR image interpretation technique,especially based on the deep learning,is one of the important directions in the remote sensing community.However,the inherent properties of SAR images have not yet been explicitly considered by the deep learning based methods,resulting in poor generalization and robustness.This thesis focuses on the research on high-resolution and single-polarized SAR images classification with Convolutional Neural Network(CNN).In this thesis,the correlation theory for building the gap between statistical representation of SAR image and feature extraction of convolution was established;with this correlation theory,statistical and robust multilayer network was presented to address the difficult problems in SAR images classification task,which captures not only the structural information but also the statistical characteristics of high-resolution SAR images.The main contributions of this thesis are given as follows.(1)The correlation theory for building the gap between statistical representation of SAR image and feature extraction of convolution was established.In this thesis,the statistical distributions and the associated parameters estimation were analyzed,and the difference between statistical representation of SAR image and feature extraction of traditional convolution was also compared.With this comparison and analysis,a bridge between the statistical representation and the feature extraction of convolution was constructed,and then the statistical representation learning theory for SAR images was established.(2)A statistical deep network based on the pixel-level statistical representation learning(PSRL)was proposed to address the problem that the pixel-level statistical characteristics of SAR images have not yet been explicitly considered by CNN.In the proposed PSRL,both the statistical representation for capturing the pixel-level statistical characteristics of SAR image and the pattern description for describing the high-level semantics were extracted on the basis of statistical representation learning theory.The presented PSRL builds a unified framework for low-level statistical representation learning and high-level pattern learning,which enhances the effectiveness of description for the random and structural SAR images.(3)A statistical deep network based on the feature-level statistical representation learning(FSRL)was presented to solve the problem that the statistical characteristics of SAR image primitives have not yet been considered by CNN.In the presented FSRL,a statistical constraint module for extracting the statistical primitives was given from the statistical point of view,which enables the feature-level statistical representation learning for SAR images based on the statistical representation learning theory.The proposed FSRL establishes a unified framework in which the representation learning and statistical analysis for SAR images are implemented.(4)Speckle-fluctuation-robust statistical deep network for SAR images classification(SRSDN)was proposed to deal with the problem that CNN is not robust to the multiplicative speckle noise.In the proposed SRSDN,both the feature extraction module and the classifier module of SRSDN were constrained with regularization,by which the parameter learning of SRSDN were forced to be invariant to the speckle noise.The proposed SRSDN improves the robustness of statistical representation learning to speckle noise,which addresses a crucial problem of CNN used for practical application.On the basis of the above research,a prototype system for automatic extraction of build-up areas from SAR images was presented,and comprehensive analysis of performance and applicable scenario of the presented methods were also given in this thesis.This prototype system integrates all the proposed methods into a unified framework,and promotes the research in this thesis to practical application.