Research On Method For Spectral-Spatial Joint Classification Of Hyperspectral Remote Sensing Image

Compared with other remote sensing images,hyperspectral images not only contain abundant ground image information,but also contain rich spectral information.It is possible to use high spectral images for higher precision image classification.However,high spectral image high dimension,high band correlation,noise and unique non-linear characteristics have brought great challenges to the analysis and classification of hyperspectral remote sensing images.Traditional hyperspectral remote sensing image classification methods are based on analysis of image pixel spectrum information only,and does not take into account the image contains abundant spatial information,such as spatial structure information,pixel location and distance information,etc.Recent research status show that joint spectral and spatial information of hyperspectral image classification method can further improve the classification precision of the image,to obtain more homogeneous area image classification figure,meet the needs of the production of the drawing.Based on the combination of spectrum and space information of hyperspectral image classification as the main line,to study how to extract and use hyperspectral image contains abundant spatial information,give priority to with spectral characteristics,spatial characteristics as supplement,the spectral characteristics and spatial characteristics(include the shape feature,texture feature)for multiple features fusion,will be merged as the input of the classifier,and the characteristics of a combination of deep learning classifier model(such as convolution neural network,deep belief networks,etc.)methods make full use of the spectral information of pixels,developed joint spectral and spatial information of hyperspectral image classification method,has obtained certain achievements.(1)For hyperspectral image classification process,the spectral characteristics of high dimension and inter-band correlation,using the statistics theory and optimization theory,for the research of hyperspectral image dimension reduction method,is proposed based on matrix related to the connection method of dimension reduction of hyperspectral images,the method in the process of building the adjacency graph is avoided using label information is difficult to determine the size of the neighborhood,and adopt more can reflect the correlation statistical characteristics of high-dimensional data correlation coefficient to measure the degree of similarity between data,build connection correlation matrix,seeking the optimal projection direction,implement the data dimension reduction.This method not only maintains the geometric structure of the data in the class,but also maximizes the distance between classes,and increases the interclass separability in the low-dimensional space.And the method does not depend on any parameter or prior knowledge.The experimental results on the standard hyperspectral image data set show that the dimensionality reduction method can increase the separability between different spectral features and is superior to other traditional methods in improving classification performance.(2)In view of the unique non-linear data structure,spatial homogeneity and heterogeneity of hyperspectral images,this paper applies statistical theory and set theory to carry out research on the extraction method of hyperspectral image features,and puts forward the extraction method of hyperspectral remote sensing image spatial spectrum features based on texture features and spatial shape features.Dimension reduction method based on matrix related to the connection dimension reduction,the principal component weight,the former is used for classification of pixel as the center of the small window to calculate the normalized central moments texture feature is obtained by mapping function,the latter section properties on the principal component weight calculation morphology,classification for pixel as the center of the small window itself and morphological properties of neighborhood pixels in section tensor form characteristics,the fusion characteristics can be more effective sample description,lay a foundation for subsequent accurate classification.Experimental results on the standard hyperspectral image data set show that the features extracted by this feature extraction method have stronger discriminant performance,low computational complexity and high classification accuracy.(3)Aiming at the problem that the existing classification methods do not make full use of the spatial information of context,and the extracted features are sensitive to the noise pixels in the neighborhood,resulting in the instability of classification accuracy,a decision fusion classification method based on joint collaborative representation(JCR)and texture features is proposed.Firstly,the joint collaborative representation model is used to decompose the sample and dictionary into multiple elements,and the corresponding collaborative representation is carried out respectively.Adaptive learning residual weight of multiple elements and linear weighting are carried out.Secondly,the multi-class SVM classifier is trained with the statistical feature calculated by the gray level co-occurrence matrix.Finally,a multiplication fusion rule is established to combine JCR with SVM.Experimental results on standard hyperspectral image datasets show that this method has better performance than other methods.(4)For most of hyperspectral image classification method to extract the shallow characteristics lead to the problem of classification accuracy is not high,on the basis of the theory of machine learning and deep learning,for the research of hyperspectral image classification method,is proposed based on kernel function of the fusion feature classification method,based on the convolution spectrum and texture profile space properties of neural network classification method,based on the empty spectral combination method of hyperspectral remote sensing image classification.The new kernel SVM,convolutional neural network and deep confidence network are adopted to build a multi-layer deep neural network model,and the resulting spatial spectrum fusion features are classified,thus improving the classification accuracy of hyperspectral images.The experimental results on the standard hyperspectral image data set show that the proposed model can achieve fast and high-precision classification effect,which is superior to some popular classification algorithms in the current hyperspectral image processing.There are 49 figures,27 tables and 165 references.