Research On Hyperspectral Data Feature Extraction Based On Manifold Embedding And High-Level Feature Analysis

Remote sensing analysis has received extensive attention in recent decades,a large number of scholars have continuously innovated and put forward many important theories in this field.Hyperspectral data has unique data structures and application scenarios,so it occupies an important position in the field of remote sensing.Hyperspectral data is a typical cube data,containing a large number of spectral and spatial features,which greatly enhances the development potential of hyperspectral data in the field of object classification and target recognition.There are much redundant information between bands,and the interference between features is great,so it will affect the accuracy of data classification and feature recognition using hyperspectral data.Taking into account the unique characteristics of hyperspectral data,the commonly used method is hyperspectral feature extraction.Pixels in hyperspectral data are distributed in manifold spaces,and these pixels actually reside on low-dimensional manifold planes.Therefore,it is feasible to transform features by doing manifold mapping and find low-dimensional expressions in manifold spaces.The purpose of this research is to minimize the correlation between features while ensuring the high-dimensional information,so that this redundant information will not interfere with the results when doing data classification and target recognition.In order to achieve this research purpose,many classical manifold learning methods are applied to the study of hyperspectral data to extract important information in lowdimensional manifold spaces.For example,a typical local linear embedding method describes the local neighborhood relationship of pixels and extracts features by constructing a nonlinear mapping.According to different sources of pixel features,these features can be divided into spectral features,spatial features and spatial-spectral features.Among them,spatial features are generally not used alone,they are always fused with spectral features to fully describe pixel information.Commonly used sample features generally include unsupervised features,semi-supervised features,supervised features,and so on.With the development of deep network models,the definition of manifold features has been fully expanded,which can effectively solve the problem that manifold structures of pixels are difficult to describe.The deep network model is effective in extracting deep features of pixels,but as a relatively new research direction,there are still many works that need to be improved and expanded.This paper focuses on the study of feature extraction of hyperspectral data,analyzes and studies various features in hyperspectral data from the perspective of manifold embedding mapping and deep feature analysis.This article mainly includes the following aspects:Firstly,in order to extract the statistical spectral features of hyperspectral pixels,to solve the problem of unclear pixel structures in manifold spaces,to describe local structures of hyperspectral data,we use statistical models to extract and analyze the statistical features of pixels.The feature extraction algorithm based on local pixel structure and statistical model is studied,and the classical high-dimensional data manifold structure is defined and expanded,which enriches the scope of application of algorithms.The rough set theory is analyzed and the neighborhood rough set that is most suitable for hyperspectral data is carefully chosen.In order to extract unsupervised statistical spectral features,a hyperspectral manifold feature extraction algorithm based on neighborhood relationships and local linear embedding is proposed.The former is used to introduce adaptive parameters in this algorithm and improve the stability of the model,the latter is used to model the local manifold structure.In order to extract semi-supervised statistical spectral features,to make full use of the advantages of label information and to extend the single neighborhood relationship,a hyperspectral semi-supervised manifold feature extraction algorithm based on multiple neighborhood relationships and patch alignment framework is proposed.In order to avoid over-fitting,a norm penalty term is introduced to limit the method complexity.Experiments prove that these two algorithms can obtain better feature extraction features from different angles and have certain advantages.Secondly,in order to extract the spatial-spectral features based on distribution matching of hyperspectral pixels and to raise the generation ability of spatial-spectral features,a hyperspectral manifold spatial-spectral feature extraction algorithm based on two-stream generative adversarial network is proposed to improve the generation ability of extracted features and to guarantee the consistency between probability distributions of pixels.Firstly,we analyze the feasibility of applying generative adversarial networks in feature extraction research,and then we analyze the feasibility of applying two-stream feature extraction models and compare them with one-stream feature extraction models.The algorithm proposed contains two streams: static stream and dynamic stream.The static features based on spectral information are constructed as the input information of the static stream,and the local change trend of the spectral reflectance in spectral spaces is used as the input information of the dynamic stream.By optimizing the structure of the generator and the discriminator,the generation ability of this model is improved.The loss function of classical generative adversarial network is improved to make it suitable for the proposed network.In these experiments,three typical hyper-parameters are analyzed and optimized,and three typical hyperspectral datasets are selected to verify the superiority of this algorithm.Experiments show that this algorithm has a high generation ability and can accurately describe the manifold structure in hyperspectral data and it has a strong anti-noise ability.Finally,in order to extract the spatial-spectral features based on pixel matching,to decrease the large differences between pixels and extracted spatial-spectral features,a hyperspectral manifold spatial-spectral feature extraction algorithm based on multi-stream variational autoencoder is proposed,using encoder-decoder models to ensure the consistency of pixels and extracted features.The variational autoencoder is selected as the basic framework to improve the generation ability of this method,maintaining the consistency between pixels and features.Firstly,we introduce the structure and loss function of variational autoencoders,and then introduce the similarities and differences between two generative models.The long short term memory network and convolutional neural network used are introduced,and the algorithm proposed is described.The spatial features of hyperspectral pixels are extracted using the local sensing code and the sequential sensing code,respectively.Then the local spatial features and sequential spatial features are extracted.The spectral feature coding is used to extract the spectral features of hyperspectral pixels,and these two spatial features are used to modify them to achieve multi-sensor data fusion and improve the discriminability of pixels.Using a variety of hyperspectral data to show that this algorithm has a better spatial-spectral feature extraction ability and it can retain the important discriminative information.Aiming at the hyperspectral data in the field of remote sensing,this paper extracts low-dimensional embedding features of high-dimensional hyperspectral pixels in manifold spaces to reduce the high correlation between features,and preserves the important information of pixels.On the basis of this,redundant information is reduced,and the ability to distinguish features is improved.In this paper,algorithms are proposed to extract different types of hyperspectral features.Firstly,to extract statistical spectral features,the statistical model and local pixel structure are merged to achieve manifold space mapping.To extract spatial-spectral features based on distribution matching,a generative model is introduced to guarantee the consistency of probability distributions,spatial features and spectral features are merged for high-precision compression of pixels.To extract spatial-spectral features based on pixel matching,a multi-stream variational autoencoder is introduced and a multi-stream feature extraction model is used to improve the adaptability.Experiments prove that the research topic in this thesis has practical significances in remote sensing data processing research.