Research On Polarimetric Synthetic Aperture Radar Images Feature Extraction And Terrain Classification

Polarimetric Synthetic Aperture Radar(Pol SAR)receives the information of the earth surface by transmitting and receiving specific electromagnetic waves.It provides rich information under all-weather and day-and-night conditions.Pol SAR image classification has a wide range of applications in the field of crop growth monitoring,natural disaster analysis,and city planning.With the development of Pol SAR,high-resolution,multi-polarimetric SAR images have abundant ground information and complex terrain scenes.It not only makes PolSAR image contain multi-type features,but also presents new challenges to the image interpretation.It requires that the feature extraction methods should be able to extract and fuse multiple types of features efficiently.Besides,with the application of Pol SAR systems,the amount of Pol SAR data is increasing rapidly.It requires that the classifier should classify effectively with a small amount of labeled and large amounts of unlabeled samples.Based on the above analysis,this dissertation starts from polarimetric and spatial information,presenting four methods for Pol SAR image feature extraction and terrain classification.The main conclusions and contributions are as follows:1.To fuse the polarimetric and spatial features,a composite kernel strategy is proposed.A series of polarimetric features and spatial features are extracted firstly.Then,according to the kernel characteristic and Mercer's theorem,the relationship between polarimetric features and spatial features is determined by a weight coefficient.This strategy not only improves the performance of a single type of features,but also concerns the complementary information between two types of features.2.For the use of spatial information in uneven and complex terrains,a multi-scale kernel method is presented.Firstly,according to the distribution that Pol SAR data satisfies Wishart distribution,the Pol SAR image is divided into multi-level superpixels from coarse to fine.Then,extract Pol SAR features at each scale and fuse them by the multi-scale kernel strategy.This method not only breaks through the limitation of single scale information on feature extraction,but also effectively aggregates the polarimetric features in different spatial scales.3.Facing the challenge of feature redundancy between joint polarimetric features,a deep learning-based mining strategy is proposed for joint polarimetric features.First,it obtains joint polarimetric features from a series of observation data and target decomposition methods.Then,it takes joint polarimetric features as the input and utilizes convolutional neural network as the feature fusing and feature mining network.This method not only makes use of the existing polarimetric feature extraction methods to interpret Pol SAR image,but also avoids the curse of redundancy of high-dimensional features,realizing the feature mining of Pol SAR information.4.Considering a large amount of unlabeled data is not used effectively,an adaptive anchor graph strategy is proposed.This method first converts the Wishart distance between the data point and its nearest anchors into the local density index.Then,the number of nearest anchors for each data point is decided by its density index.After that,the anchor graph is built by the proposed strategy.This method not only breaks the restrictions on the image,terrain type,and data local density,but also makes full use of the unlabeled data.The proposed methods in this dissertation have been verified by two real Pol SAR datasets.The results show that the proposed methods can effectively solve the main issues in Pol SAR image feature representation and terrain classification,achieving highaccuracy.