The Research On Feature Extraction Methods Of Hyperspectral Remote Sensing Images

With the development of hyperspectral remote sensing technology,hyperspectral remote sensing images have been widely used in the fields of precision agriculture,enemy reconnaissance,geological exploration and environmental protection etc.But the classification performance of ground objects is not ideal compared with the traditional multi-spectral remote sensing images,the main reasons lie in that the hyperspectral remote sensing images have the characteristics of multi-band,high spatial resolution and information redundancy that caused by high-dimensional data,moreover,high-dimensional data can even lead to Hughes phenomenon.Based on these characteristics,feature extraction of hyperspectral remote sensing images has naturally become one of the main research hotspots.Higher spatial resolution increases the complexity of terrain types and brings great challenges to feature extraction.And traditional image processing technology can hardly meet the needs of practical application because of the high dimensionality of hyperspectral remote sensing images data,the strong correlation between spectral bands,the peculiarity of multi-dimensional representation and the integration of atlas.Therefore,it is very important for us to study the feature extraction methods of hyperspectral remote sensing images.Based on the theory of machine learning and optimum mathematical modeling,aiming at the shortcomings of existed feature extraction methods of hyperspectral remote sensing images,this research paper improves and innovates the feature extraction technology of spectral information and spatial information.Meanwhile,it also extends the research scope from spectral dimensionality to spectral-spatial dimensionality.The main research contents and important conclusions of this paper can be summarized as follows:(1)Aiming at feature extraction of hyperspectral remote sensing images in spectral dimensionality,a novel feature extraction method based on Maclaurin function curve fitting feature extraction(MFCF-FE)is proposed.The proposed MFCF-FE method mainly solves the feature extraction problem in spectral dimensionality of hyperspectral remote sensing images with limited training samples,and fully mines the geometric features that reflected by spectral response curves of each pixel that located on each spectral band.The experimental results show that this feature extraction method can improve the classification accuracy and show good robustness.(2)Aiming at the problem of feature extraction in spectral dimensionality of hyperspectral remote sensing images and finding the optimal discriminant vector in transformed low-dimensional space,we propose a feature extraction method based on geometric mean feature space discriminant analysis(GmFSDA).The proposed method mainly solves the feature extraction of hyperspectral remote sensing images under the condition of limited training samples.Firstly,we obtain the first projection transformation matrix by optimizing the geometric dispersion matrix between spectral bands,and meanwhile,the original data is transformed into a new feature space.The main purpose of this transformation is to enhance the difference between spectral bands.Secondly,an optimization model is established to find the optimal discriminant vector that is called the second projection matrix in the transformed feature space.Whose aim is to make the difference of the same kind of objects as small as possible,and the difference between different types of objects is as large as possible.The proposed GmFSDA feature extraction method realizes the transformation of data from high-dimensional feature space to low-dimensional feature space.The experimental results show that the GmFSDA feature extraction method can better reveal the inherent geometric structure characteristics of hyperspectral remote sensing images,and can achieve higher classification accuracy than the other feature extraction methods.(3)In order to utilize the spatial structure features of hyperspectral remote sensing images,a new feature extraction method based on harmonic mean spectral-spatial filter(SSF_HM)discriminant analysis is proposed.In this method,we use area median filter(AMF)to characterize the spatial structure of hyperspectral remote sensing images and extract the spatial features by filtering thep_i principal components that obtained by principal component analysis(PCA).Then fusing the original spectral features and extracted spatial filter features to form fusion feature matrix,after that,we estalibish an optimization model based on harmonic mean spectral-spatial filter inter-class and intra-class scatter matrix and solve it by using Fisher discriminant analysis.The feature extraction method not only combines spectral dimensionality and spatial dimensionality,but also realizes the transformation from spectral dimensional space to spectral-spatial filter dimensional space.The experimental results show that the proposed SSF_HM feature extraction method can enhance the separation of data and obtain higher classification accuracy.(4)In order to further utilize the spectral and spatial features of hyperspectral remote sensing images,we use Gabor transformation to extract the Gabor features,whose purpose is to obtain more precise and effective spatial features.Aiming at mining the Gabor spatial features of hyperspectral remote sensing images from different scales and directions by using Gabor filter,a feature extraction method based on spectral-Gabor spatial discriminant analysis(SGDA)with limited training samples is proposed.The proposed method uses Gabor filter to describe the spatial structure of hyperspectral remote sensing images and extract Gabor spatial features.Meanwhile,we solve the optimization model about intra-class and inter-class scatter matrix by singular value decomposition(SVD)and feature space transformation(FST).The proposed SGDA feature extraction method can fully demonstrate the strong representation of hyperspectral remote sensing data in low-dimensional feature space,and transform the space from high-dimensional feature space to low-dimensional feature space.The experimental results show that the SGDA feature extraction method is indeed effective and feasible.The main conclusions are as follows:in the case of limited training samples,from the point of view of curve fitting,a mathematical fitting model is established in spectral dimensionality to mine the geometric structure features that reflected by spectral response curves of band pixels.And then we propose a feature extraction method based on Maclaurin function curve fitting(MFCF).In order to find the solution vector of the optimal model in low-dimensional space,a feature extraction method based on geometric mean feature space discriminant analysis(GmFSDA)for limited training samples is proposed,which realizes the conversion of space from high-dimensional feature space to low-dimensional feature space.In the spectral-spatial dimensionality,the spatial structure of hyperspectral remote sensing images is characterized by means of area median filter(AMF).Thus,a feature extraction method based on harmonic mean spectral-spatial filter(SSF_HM)discriminant analysis is proposed and realizes the combination of spectral dimensionality and spatial dimensionality.In order to extract more precise and effective spatial features,we use Gabor transformation to obtain Gabor space features from different scales and directions of hyperspectral remote sensing images.Therefore,we propose a feature extraction method based on spectral-Gabor spatial discriminant analysis(SGDA)with limited training samples,this method realizes the combination of spectral dimensionality and Gabor spatial dimensionality effectively.And the combination enlarges the range of feature extraction and provides a research basis for multi-features fusion in the future.