Research On Hyperspectral Image Classification Algorithm Based On Spatial Spectral Combination

Deep learning has been widely used in hyperspectral image classification tasks because of its ability to autonomously mine and characterize the characteristics of hyperspectral images.However,classification algorithms based on deep learning tasks still have problems such as data redundancy,insufficient fine feature extraction,and loss of global and local spectral information,and many methods only consider spectral information or spatial information when classifying hyperspectral images.Can’t make full use of image features.Therefore,after analyzing and summarizing the status quo of hyperspectral image classification,this article proposes two hyperspectral image classification algorithms based on space spectrum combination,mainly from dynamic convolution and graph convolution.To solve the current problems encountered in joint hyperspectral image classification based on space spectrum.In order to solve the problems of data redundancy and insufficient extraction of interaction features between fine features and space spectrum in hyperspectral image classification task,etc.A hyperspectral image classification method combining dynamic convolution and triple attention mechanism is proposed.Firstly,the data is preprocessed,and the dimension is reduced by principal component analysis to remove redundant information.Then,the feature extraction module of residual network combined with dynamic convolution improvement was constructed,and dynamic convolution adaptively adjusted convolution parameters to extract in-depth and refined features.After dynamic convolution layer,triple attention mechanism is introduced to interact with each other across dimensions to obtain more discernible hyperspectral space spectrum information.This method efficiently captures the fine features and cross-latitude interaction space spectrum information.Classification experiments were carried out on three public data sets of Pavia University,Kennedy Space Center and Salinas respectively,and the results show that the proposed method enhances the classification performance of network models and effectively improves the classification accuracy.Secondly,a model based on graph convolution network is proposed for hyperspectral image classification in view of the large amount of information in hyperspectral images and the difficulty in obtaining global and local marginal information.Firstly,the data is processed into superpixel nodes by using the superpixel algorithm,and the generated graph nodes are more suitable for graph convolutional network convolution operations.At the same time,graph convolution can supplement the information of neighbor nodes as the information of the current node to obtain marginal information and clearer boundary features,and the hyperspectral features obtained are more complete than the single individual features.Multi-scale operation is introduced into graph convolutional network,which fully mines local features of hyperspectral images from different scales and extracts more abundant space spectrum joint information.The results of three open data sets show that this method achieves better classification effect and higher classification accuracy.