Research On SAR Image Object/Land Fine Interpretation Methods

With all-day and all-weather image acquisition capability,relatively strong penetration capability,as well as polarimetry characteristics,Synthetic Aperture Radar(SAR)has found various applications in military and civil community.SAR image interpretation is to acquire key information via the interaction between objects and land covers.Recently,with the rapid increase of SAR images and the wide application of machine learning methods represented by Convolutional Neural Networks(CNNs),a single technology or method is difficult to satisfy the requirement of SAR image fine interpretation,which tends to the combination of traditional models and deep CNNs.Based on multiple machine learning methods,such as CNNs and Gaussian Process Re-gression(GPR),this thesis aims to research fine interpretation of earth observations for SAR images from the aspects of object detection,polarimetry,detecting wavebands,as well as dataset compilation and quality evaluation.Main contents of this thesis include:object-oriented fine interpretation based on SAR ship detection,polarizations-oriented fine interpretation based on physical scattering characteristics,waveband-oriented fine interpretation based on physical scattering characteristics,and categories-oriented land cover fine interpretation based on dataset compilation and quality evaluation.Firstly,this thesis realizes high precision SAR ship detection in open sea and nearshore areas by using a revised multi-scale CNN.In order to further attain more accurate detections,reduce missing detections in densely clustered area and false alarms on land,this thesis presents a method combining cascade CNN and Pulse Cosine Transform(PCT)based visual attention method.Aiming to achieve SAR ship candidates,the former analyzes SAR images in the space domain.On the contrary,the latter analyzes SAR images in the frequency domain in adaptive local regions.The superiority of the proposed method has been demonstrated on European Space Agency(ESA)'s Sentinel-1 dataset and the Chinese Gaofen-3(GF-3)dataset.Secondly,in order to extract physical signatures from single-and dual-polarimetric SAR images and further explore potentials on SAR images with multiple polarizations,this thesis proposes a contrastive-regulated CNN in the complex domain,directly learning physically interpretable deep model from complex Sinclair scattering matrices.The input,output,and parameters are all complex values,and the computations are in the complex domain.The loss function is composed of a basic loss and a contrastive regularization term.This study specifically focuses on multiple scattering,volume scattering,and surface scattering.With the assistance of Cloude's polarimetric decomposition,this thesis generates ground-truth of physical scattering types.Experiments on DLR's F-SAR data demonstrate the effectiveness and generality of this method.Thirdly,this thesis realizes wavebands-oriented fine interpretation for SAR image scatter-ing characteristics based on GPs.Based on Gussian hypothesis conditions,this thesis proposes a kernel-and expectation-based Gaussian process regression method to learn some polarimetric parameters,such as polarimetric entropy and polarimetric alpha angle.Then,the physical scat-tering signatures for earth observations can be obtained with the help of H-?division plane.Gaussian processes have the capability of realizing high quality prediction and their correspond-ing uncertainty levels.The kernel trick and expectation operations could alleviate geometric distortions and calibration discrepancies beween images acquired with different wavebands.Experiments on X-band Terra SAR-X images,C-band Sentinel-1 images,and L-band F-SAR images demonstrate the effectiveness and robustness of this method.Finally,this thesis contributes a large-scale SAR database with relatively low resolution.The images are acquired by Sentinel-1 in 21 Chinese major cities.Based on a pre-defined hier-archical annotation system,the dataset is divided into ten different categories,mainly including building types.The annotation is guaranteed by the transition from optical annotation in Google Earth to SAR annotation by Sentinel-1 Application Platform(SNAP).The database provides33,358 image patches with 100_×100 pixels,each patch is with VH and VV polarizations and four different data formats(including original 32-bit data,UINT8 data,radiometric-calibration data,and pseudo-color data).The quality of this dataset is evaluated via fast compression distance based manifold visualization method and multiple image classification methods.