PolSAR Image Classification And Ship Detection Based On Deep Learning

Polarimetric Synthetic Aperture Radar,as an advanced means of obtaining remote sensing information,can acquire the scattering information of different objects under different polarization receiving and transmitting modes all day,all weather and far distance.The strong information acquisition capability of PoISAR significantly enhances its ability to detect and classify objects.PolSAR has broad application prospects in civil environment monitoring,crop yield estimation,resource exploration and other civil and battlefield reconnaissance and situation monitoring.Currently,the performance of object detection and classification algorithms for PolSAR images mainly depends on the model that extracts object feature designed by using prior information.However,inaccurate prior information and models can easily affect the performance of object detection and classification.The deep learning method can not only adaptively extract low-dimensional features such as two-dimensional spatial feature of the object,but also fully exploit the high-dimensional feature of the data.Therefore,the research of PolSAR image classification and object detection based on deep learning has important theoretical significance and application value.Based on the National Natural Science Foundation of China,the New Century Excellent Talent Support Program of the Ministry of Education,and the horizontal project,this thesis focuses on the PolSAR classification and object detection based on deep learning and obtains good results.The main contributions of this thesis are listed as follows:1.The basic theory of polarization is elaborated.The characteristics and forms of PolSAR data are analyzed in detail from the aspects of polarization characterization of electromagnetic wave,polarization scattering mechanism and polarization decomposition theory,which lays a theoretical foundation for the following PolSAR image classification and ship detection based on deep learning.2.Aiming at the problem that the convolutional neural network classifies the PolSAR image with low computational efficiency and can not achieve end-to-end classification,a method based on SegNet is proposed for PolSAR image classification.This method has the ability to accept any size of PolSAR image as input and obtain the end-to-end pixel-level classification results.Firstly,the Deep Convolutional Neural Network(DCNN)encodes the features of the object in the PolSAR image.Then the DCNN with the anti-pooling layer is used to decode the encoded result and output the score of the feature classification,thus acquiring the classification map.The effectiveness of the proposed method is verified by the classification results of the measured polarization data.3.Aiming at the problem that the current object detection algorithm based on Constant False-Alarm Rate(CFAR)is not accurate and easy to cause false alarm,an improved PolSAR ship detection method based on Faster R-CNN is proposed.In this method,DCNN is used to extract the hierarchical features of ship targets automatically,and multi-scale features are used to generate candidate frames adapted to ship targets of different scales.Then Faster R-CNN framework is used to realize fast and robust detection of multi-scale ship targets.At the same time,this thesis also uses polarization data generated by AIRSAR and UAVSAR platform to establish the database of ship,sea surface and coast samples,designs a classifier based on DCNN to achieve fast and accurate land-sea segmentation,and restrains the impact of coast on ship target detection in the scene.The effectiveness of PolSAR ship detection method based on improved Faster R-CNN is verified by detection results based on measured polarization data.