Hyperspectral Image Classification Algorithm Based On Singular Value Decomposition And Spatial-spectral Synergistic Enhancement

Hyperspectral image classification is an important branch of remote sensing image processing.Hyperspectral image is high-dimensional image data composed of spectral information of hundreds of bands and two-dimensional spatial information corresponding to each spectrum.The spectral information contained in the hyperspectral image data is helpful for us to distinguish the land cover,and the fine spatial resolution provides a wealth of spatial structure information.Therefore,hyperspectral images have played an important role in geological disaster monitoring,mineral analysis,agricultural experiments,and marine environmental surveys.Therefore,the correct classification of hyperspectral images is very important.Traditional classification methods treat hyperspectral images as ordinary images.The shallow classification model proposed by them is often difficult to extract the essential features of the image,which affects the classification accuracy.After the advent of deep learning,it has been widely used in image processing related fields because of its ability to extract the deep characteristics of input data and the ability of nonlinear mapping.However,because hyperspectral images have serious problems with the same spectra and different spectra,the application of deep learning to hyperspectral images faces serious challenges.This paper proposes our solutions to the problems existing in the classification of hyperspectral images.The specific research contents are as follows:(1)In view of the characteristics that hyperspectral image data is three-dimensional data containing spectral dimensions and corresponding two-dimensional spatial information,because the two-dimensional convolutional neural network cannot extract better identification feature maps from the spectral dimensions,the three-dimensional volume The problem that the product neural network has worse performance in many spectral bands with similar texture categories,we use densely connected deep three-dimensional convolutional neural network to extract the spectral information and spatial information of the hyperspectral image,and add attention on this basis The force mechanism module further enhances the feature information and greatly improves the classification accuracy.(2)Aiming at the problem of redundant information and noisy information in hyperspectral image data,we combined the traditional feature extraction method and the feature extraction method of convolutional neural network to design a convolution-like feature extraction network based on singular value decomposition,The feature extraction network starts from its own data and obtains the singular value matrix through the singular value decomposition method.The singular value vector is transformed into a"convolution kernel",and this"convolution kernel" is used with its own calculations to achieve similar volumes.The effect of the product.Make the extracted features more discriminative,thereby improving the classification effect of hyperspectral images.(3)Aiming at the problem of space spectrum feature fusion of hyperspectral images,we designed a two-stage dual-branch hyperspectral space-spectrum feature fusion classification network,which is the first-stage feature extraction network and the second-stage dual-branch network.Classification network.Different improved attention mechanism modules are used in the two branches to enhance the spectral and spatial features.Finally,the enhanced features are fused and classified in a cascade manner,which makes the classification accuracy rate higher.