Research On Classification Of Hyperspectral Remote Sensing Images Based On Graph Neural Network

Due to the development of imaging spectrometers,remote sensing detection has entered the hyperspectral stage.Hyperspectral imaging organically combines original imaging technology and spectroscopy technology.It uses imaging technology to gather spatial information of the earth's surface and uses spectral technology to decompose the total radiation in each pixel into radiation spectra of different bands.The hyperspectral remote sensing image obtained is no longer a traditional two-dimensional image,but a three-dimensional spectral data cube.And the spectral data cube contains not only the image information of each band but also the spectral information of a single pixel,which realizes the true “unification of image and spectral”.Hyperspectral image classification is a key technology in the field of hyperspectral remote sensing,which is of great significance to meteorology,environment,agriculture,and military.Hyperspectral image classification is for each hyperspectral pixel which is a vector formed by the radiation of different bands,and each vector corresponds to a spectral curve.Surface materials exhibit different radiation in different wavelength bands due to different material compositions and lighting conditions.According to the different radiation,determine the material represented by the hyperspectral pixel and assign a material label to the hyperspectral pixel,which is the main content of the hyperspectral image classification.For the classification task of hyperspectral remote sensing images,statistical methods and machine learning methods have been applied successively,but the accuracy that can be achieved is limited.With the development of deep learning technology,researchers have introduced deep learning methods into the field of hyperspectral image classification,and have achieved great success.Among the deep learning methods,the convolutional neural network is the most widely used method and has become the mainstream method of hyperspectral image classification.Hyperspectral images are typical Euclidean data,and researchers have been using th methods which are used to process Euclidean data to process hyperspectral images,such as convolutional neural networks,but this ignores the relationship between pixels in irregular regions in hyperspectral images.Due to the imaging mechanism,hyperspectral images are Euclidean data,but the real situation of the ground surface may not be expressed by regular-shaped grid sampling.The relationship between pixels in non-Euclidean connections is also worthy of attention.In view of the fact that the existing deep learning methods applied to hyperspectral image classification cannot well represent the relationship between nodes(pixels),this paper applies graph neural networks to classify hyperspectral remote sensing images.This paper proposes three different graph neural networks for hyperspectral remote sensing image classification.The proposed three graph neural networks give three different solutions on how to convert hyperspectral images into graph structure data,how to represent node relationships to better extract features.(1)We propose a novel network named Superpixel-based Graph Attention Network for hyperspectral image classification.The network applies a superpixel segmentation algorithm and then constructs a graph structure according to the spatial connection relationship of superpixels.Using this method,the number of graph nodes is small,which means the network has the characteristics of small computing resource occupation and fast network running speed.After obtaining the graph structure composed of superpixel nodes,a multi-head attention mechanism is applied to characterize the relationships between nodes.The multi-head attention mechanism can comprehensively represent the relationship between nodes and themselves,nodes and neighbors from multiple dimensions,and perform feature fusion and extraction.The network is compared with other methods of many types on three hyperspectral benchmark datasets,and the experimental results show that the classification ability of the network is excellent.(2)We propose a novel network named Global Random Graph Convolution Network for hyperspectral image classification.The network applies a global random graph construction scheme in which the graph structure is constructed by random sampling from the labeled samples of each class and combining them in sequence.Constructed by this scheme,the graph structure is small in scale,which can save computing resources;huge in number,which solves the problem of insufficient samples in hyperspectral image classification to a certain extent;diverse in combination,which makes the trained network is more robust.The network also designs parallel feature extraction networks and cascaded graph convolution modules.The feature extraction network takes into account both the spectral dimension and the spatial dimension to extract features;the graph convolution module uses a neural network with trainable parameters instead of manual rules to determine the adjacency matrix,which can better mine the relationship between nodes.The network is compared with other methods of many types on three hyperspectral benchmark datasets,and the experimental results show that the classification ability of the network has reached an advanced level.(3)We propose a novel network named Hierarchical Graph Transformer for hyperspectral image classification.This network introduces the graph transformer into the field of hyperspectral image classification for the first time and applies a local random graph construction scheme with three different strategies.The construction of the local random graph transforms the hyperspectral image classification problem into a ”graphlevel” classification task,and at the same time greatly increases the number of trainable samples,which solves the problem of insufficient samples to a certain extent.The network design a orientation-based edge feature for position encoding in transformer architecture.The encoding method is to use the extracted edge features to modulate the subsequent attention coefficient matrix,which provides a position encoding solution for applying the graph transformer network for hyperspectral image classification.The network also designs a hierarchical training mechanism in which the network is trained using a hierarchical structure of ”inner graph-outer graph”.And the hierarchical mechanism can greatly reduce the time required to train the network while making better use of spatial information and improving classification accuracy.The network is compared with other methods of many types on three hyperspectral benchmark datasets,and the experimental results show that the classification ability of the network has reached an advanced level.