Research On Target Detection Methods For PolSAR

Polarimetric synthetic aperture radar(PolSAR)is a new radar system based on the traditional synthetic aperture radar(SAR)system.It not only has the advantages of full-time,all weather,long range,wide-swath and high-resolution imaging,but also can make full use of the polarimetric information of electromagnetic waves to describe the target physical scattering process.Therefore,it plays a very important role in the fields of target detection,resource exploration,topographic mapping,sea surface environmental monitoring,battlefield monitoring and reconnaissance,etc.However,the complex representation of PolSAR data,the complicated observation scene,and the radio frequency interference(RFI)in the same frequency band acquried by the PolSAR increase the difficulty of target detection for PolSAR data.The complexity of PolSAR data representation greatly enhances the difficulty to model the statistical distribution of PolSAR data and accurately estimate the model parameters,which is a hot topic in the field of target detection for PolSAR data.Meanwhile,in the complicated observation scenes,the scattering characteristics of targets can be easily interfered.Thus,how to design an effective feature extraction method and mine the correlation of various features of targets is also a key issue in the field of the target detection for PolSAR data.Simultaneously,the existence of RFI also degrades the performance of PolSAR imaging and subsequent target detection.The detection and and mitigation of interference signals is a hot issue in the field of radar signal processing.Therefore,the research on the target detection method for PolSAR data is of important theoretical and practical significance.In order to improve the performance of target detection,this thesis focuses on the research of RFI mitigation,statistical distribution model representation,parameter estimation,feature extraction and target detection methods for PolSAR images.The main contributions are listed as follows:1.Based on the working principle of PolSAR,characterization methods of for PolSAR imaging results are studied,thus the statistical characteristics of the PolSAR imaging results and the scattering characteristics of targets are analyzed.For the sake of establish a statistical model of PolSAR data and analyze its statistical properties,the double product stochastic model is introduced.Then,Mellin transform and the Mellin kind statistics are utilized to analyze the statistical characteristics.Meanwhile,the scattering characteristics of targets are analyzed by studying typical polarimetric target decomposition methods.Then,from the aspects of RFI representation and RFI characteristics analysis,the influence of interference on PolSAR data in different characteristic domains is qualitatively analyzed,which lays a foundation for the research of subsequent RFI mitigation methods and target detection methods for PolSAR data.2.Aiming at the problem that RFI seriously affects the imaging quality of PolSAR and the performance of target detection for PolSAR data,this dissertation proposes a deep convolutional neural network(DCNN)based RFI detection method and a deep residual network(ResNet)based interference mitigation method.Firstly,the methodologies of DCNN is introduced in detail.According to the characteristic differences between RFI and target echo signal in time-frequency domain,a DCNN-based RFI detection algorithm is proposed.The proposed algorithm transforms the RFI detection problem into a binary classification problem using classical visual geometry group(VGG)to extract the characteristics of the RFI in time-frequency domain and detect the echoes containing narrow-band interference(NBI)and wide-band interference(WBI).Then,a RFI detection algorithm based on CNN and a RFI mitigation algorithm based on ResNet are proposed.According to the characteristic difference between the RFI and the target echo signal in timefrequency domain,the proposed method uses ResNet and skip connection structure to extract the characteristics of target echo signal,thus realizing the reconstruction and removal of RFI.The proposed algorithm can not only improve the accuracy of NBI and WBI mitigation,but also speed up the process of interference mitigation.Finally,the NBI and WBI detection and mitigation results on the simulated and measured airborne and spaceborne SAR data verify the effectiveness of the deep learning-based interference detection and mitigation algorithm.3.To solve the problem that Gaussian model cannot accurately represent the statistical distribution of PolSAR data and may lead to poor ship target detection performance,this dissertation proposes a non-Gaussian distributed algorithm to improve the ship detection performance for PolSAR data.Firstly,the non-Gaussian K-Wishart model and its ability to characterize PolSAR data is studied.Then,the sampled PolSAR data is clustered using the modified expectation maximization(EM)algorithm.For the K-Wishart distribution model is complex and the parameter estimation is difficult,the Mellin transform and Mellin kind statistics are introduced to estimate the parameters of the proposed model.The number of clusters in the EM algorithm can be determined by goodness-of-fit test.After obtaining the unsupervised clustering results of PolSAR data,SPAN,together with the difference of polarization scattering characteristics between the ship targets and sea background,are utilized to realize ship detection in the unlabeled clustering results.Finally,multiple datasets acquired by different PolSAR platforms are used to verify the effectiveness of the proposed ship detection algorithm.4.Aiming at the problem of poor performance of ship target detection for PolSAR data in complex observation environments,this dissertation proposes a ship detection algorithm based on modified Faster R-CNN.Strong clutter in complex observation environments will have strong scattering characteristics and fail to detect the weak ship targets.Meanwhile,coasts have similar features to ship targets in low-resolution PolSAR images,which reduces the detection accuracy of weak ship targets.The traditional Faster R-CNN combined with DCNN can capture the difference of targets in high-dimensional feature domain.It can accurately detect various targets in optical images,while the traditional Faster R-CNN has poor detection ability for small targets.To solve the problems above,this dissertation firstly increases the sample diversity and suppresses the speckle noise and clutter by preprocessing the PolSAR images.Then,DCNN is used to develop a sea-land segmentation network to suppress the impact of coasts on ship detection.Next,Faster R-CNN is modified to produce multi-scale proposals to improve its performance on small ship target detection.Finally,the ship detection results are fused by combining the geometric relationship of samples and nonmaximum suppression method to obtain final results.The effectiveness of the proposed ship detection method is verified by multiple PolSAR datasets acquired by AIRSAR and UAVSAR.