Research On Multiple Feature Fusion And Classification Of SAR Images

Synthetic Aperture Radar(SAR)is an active imaging radar system with high resolution in all-time and all-weather conditions,which has been widely used in economic,military and other fields.However,the existence of inherent speckle noise in SAR image seriously affects the image quality,and brings great challenges in the understanding and interpretation of SAR image.Land cover classification of SAR images is an important part of the understanding of SAR images.Although abundant research achievements have been made in this field,there are still some problems such as little prior knowledge,multi-feature fusion and optimization,effective use of spatial structure information,multi-feature linear inseparability and adaptive classification modeling.This thesis aims at solving the above problems,and it integrates the tensors theory,the sparse representation of kernels and the adaptive kernel technology to fully mine spatial structure information of SAR image,using the correlation within the neighborhood and diversity in non-neighborhood.The algorithm of multifeature extraction,fusion and classification for SAR image are studied deeply,and the following research results are obtained:(1)A semi-supervised classification algorithm based on tensor decomposition and clustering is proposed.The feature-based SAR image classification method has the problem of missing spatial structure information.How to retain the structural information of SAR image,improve the feature-based recognition ability and reduce the impact of noise is still a challenge in this field.For this reason,this paper first uses clustering method to keep the non-local information in the image and tensor,and can retain the spatial structure information.Secondly,block clustering algorithm is used to generate multiple high-order cluster tensors of multi-feature SAR images,forming a multi-manifold structure,and retaining the local and non-local spatial structure information inherent in the image.Then,by considering the local structure,labeled and unlabeled information,the improved discriminant analysis and tensor decomposition theory are combined to generate multiple new projection directions of cluster tensors,which enhances the in-class compactness and inter-class separability,and improves the ability of feature-based recognition.Finally,the classification of SAR images is realized based on support vector machine and verified by experiments.(2)A new SAR image classification algorithm based on multi-feature and adaptive kernel function combination is proposed.The algorithm classifies SAR images by studying the adaptive composite kernel and composite weight fusion strategy.First,the gray-level co-occurrence matrix,wavelet energy and attribute profiles are extracted from SAR images to construct three complementary 3D feature tensors.Then,according to the Gamma distribution of SAR image and the negative logarithm likelihood values,the three dimensional feature tensor of SAR image is divided into the three dimensional feaure blocks,then put forward the adaptive composite kernel strategy to mine the context spatial information of within each feaure block,and according to the spatial structure information of each feature block,the weights in the new composite kernel are automatically determined.Finally,the final classification results are obtained by constructing the decision fusion strategy of composite weights,and the effectiveness of the method is verified by experiments on synthetic and real SAR images.(3)A SAR image classification algorithm based on multi-feature non-local dynamic kernel sparse representation is proposed.In order to solve the problem of linear inseparability caused by multiple features of SAR images,this method firstly extracts multiple features of SAR images,and constructs the kernel space of different tensor features through mapping and kernel function.Secondly,in order to make full use of the local similarity and the difference of non-local information about different types of feature space structure,and reduce the interference of non-local information,tensor projection is carried out on the kernel space and a multi-feature non-local dynamic kernel sparse representation model is constructed based on it.Then,based on the multi-feature non-local dynamic kernel sparse representation residual,the classifier is constructed to realize the classification of SAR images.Finally,experiments are carried on the simulation and real SAR images to compare and verify.(4)A SAR image classification algorithm based on attention-aided stacked sparse autoencoder is proposed.In order to solve the problems of label sample limited,local spatial information missing and recognition accuracy on SAR image,this method firstly extracts three different features from the original SAR image,these construct the spectral texture information of the attention module.Through the dimensionality reduction and Gabor filtering on these texture features,the spatial information of the attention module are formed.Secondly,for solving the problem of label sample lack,according to the similarity and distance of real samples virtual training samples are generated,and stacked sparse autoencoder is pre-trained with real samples and virtual samples.Then,logistic regression combined with stacked sparse autoencoder is used to fine-tune the whole deep network,and they realize SAR image classification.Finally,experiments are carried on the real SAR images to compare and verify.