Research On Spectral–spatial Joint Hyperspectral Image Classification Technology Based On Deep Learning

In recent years,hyperspectral imaging technology has made rapid development,which makes hyperspectral image classification technology gradually become one of the hot research directions in this field.Hyperspectral images not only have fine spectral characteristics,but also accurately reflect the spatial distribution characteristics of ground object categories.It provides the main classification basis for ground object classification and has very important application value in target classification and recognition.In previous studies,researchers mainly focused on the extraction of spectral feature information,ignoring the important impact of the use of spatial information on the improvement of image classification performance.Therefore,hyperspectral feature extraction combined with spatial information and spectral information provides a new technical idea for high-precision image classification.In view of the above problems,based on hybrid convolutional neural network,this paper aiming at optimizing feature extraction constructs two hyperspectral image classification models combined with residual network and attention mechanism respectively from the perspective of spatial-spectral joint.Through experimental verification and comparative analysis of the classification effects of various algorithms,the experimental results confirm that the two algorithm models proposed in this paper have significantly improved the classification accuracy of hyperspectral images.The main research contents are as follows:1.This paper briefly states the research background and application value of this topic,introduces the development process and research status of hyperspectral image classification methods,and expounds the detailed concept of remote sensing hyperspectral image classification combined with its data representation characteristics,and summarizes the evaluation criteria of classification performance and several classical hyperspectral image classification comparison algorithms.2.In view of the large noise interference of spectral images in the classification of hyperspectral images,the problems of insufficient feature distinguishability and low classification accuracy existed in relying solely on spectral information to achieve classification,a classification method based on space spectrum information fusion based on hybrid residual convolutional network is proposed.First,we discuss the influence of the residual network on the gradient in the deep convolutional neural network and the reliability of the effective fusion of spatial spectrum information.Then design the 3DCNN and 2DCNN mixed residual convolutional neural network to extract the spectral features and spatial features from the hyperspectral data,respectively.The residual network is used to compensate for the gradient decline and disappearance,which improves the classification performance,weakens the over-fitting situation in the process of network training,and improves the generalization ability of the model.3.Aiming at the problem that the low reliability of neighborhood information selection in feature extraction by convolution neural network which easily leads to the difficulty of accurate classification.This paper proposed a spatial-spectral joint classification algorithm combined hybrid convolutional neural network with attention mechanism.The framework takes as the main way to improve the effective utilization of spatial information in the classification process,starting from the neighborhood selection of image spectral dimension and spatial dimension information,and applying hybrid 3D-2DCNN convolutional neural network for feature extraction and classification.It avoids the influence of the introduction of inappropriate neighborhood information on the classification accuracy,improves the classification performance of hyperspectral images,and enhances the anti-noise interference ability in the classification process of hyperspectral images.