Study On Ship Detection Methods For Polarimetric SAR Image

The monitoring and detection of marine vessels is the key to maintaining world security,which is also an urgent issue that coastal countries must take effective measures to deal with.The polarimetric synthetic aperture radar（SAR）is a microwave imaging system that uses active sensors to detect ground targets.Despite of advantages such as multi-polarization beam scanning,high resolution,strong anti-interference ability,and strong penetration,the polarimetric SAR can provide abundant backscattering information to obtain ccurate inversion of target parameters.Therefore,the ship target detection based on polarimetric SAR images has been one of the hot research directions.However,it is unavoidable to face problems,including the azimuth ambiguity of ship targets,redundant polarimetric features,and complex representation of polarimetric SAR data,which pose challenges to target detection with polarimetric SAR images.To address the above issues and improve detection performance,the research of this thesis can be summarized as follows.（1）To reduce false alarms caused by the azimuth ambiguities of ship targets,an azimuth ambiguity removal method based on the geometric perturbation-polarimetric notch filter（GP-PNF）is proposed.This method combines the third eigenvalue and the new GP-PNF（denoted as NPNF）.To be specific,the third eigenvalueλ₃ is obtained by the conducting eigendecomposition with the polarimetric covariance matrix of the polarimetric SAR image.Based on the λ₃,a new eigenvector can be obtained.Besides,the target energy can be obtained with the NPNF.Therefore,the new eigenvector can be used to eliminate the azimuth ambiguity under the framework of GP-PNF.Finally,experiments are carried out on three sets of polarimetric SAR data sets.The experimental results validate the ship detection performance and the ability to remove azimuth ambiguity of the proposed method.（2）Aiming at the redundancy of polarization features,a random forest based ship detection method is proposed for polarimetric SAR images.Firstly,the polarimetric decomposition algorithms are used to extract multiple features to construct the initial feature set.With respect to the feature selection,the fast correlation based filter（FCBF）feature selection algorithm is used to calculate the correlation between features to remove the redundant features.Finally,the binary classification problem of ship and sea clutter is realized with the random forest algorithm,which improves the classification efficiency.Verified by the polarimetric SAR image of the sea area near Kojimawan Bay,Japan,the proposed method can effectively improve the detection rate of ship targets.（3）To overcome the shortage that the traditional real neural network cannot use the phase information of polarimetric SAR images,the complex-valued convolutional neural network（CV-CNN）is used to detect ships from the polarimetric SAR images.To fully utilize the phase information of the polarimetric data,the complex backpropagation method based on stochastic gradient descent principle is derived for the training of CV-CNN.Finally,the detection performance of the CV-CNN is verified with real-measured polarimetric SAR images.