Hyperspectral Remote Sensing Image Classification Based On Graph Convolutional Network

Hyperspectral remote sensing image is a three-dimensional spectral cube,which contains complete image information in each band.Viewed from a single pixel,the three-dimensional cube contains spectral information for each pixel.Therefore,the difficulty of hyperspectral remote sensing image classification lies in the fact that the number of spectral channels is much more than that of ordinary images,and it has rich spatial features,texture features and band features.Therefore,the problem of "same object with different spectrum and foreign object with the same spectrum" also exists.The neural network model formed by combining the data structure formed by graph representation learning with the convolutional neural network of Euclidean space is called graph neural network.Graph representation learning can build the dependence relationship between nodes in non-Euclidean space.In terms of remote sensing images,graph representation learning can extract spectral features from data,thus solving the problem of remote sensing image classification.Graph neural network has two branches.Graph convolutional network is a branch of graph neural network in frequency domain.Hyperspectral remote sensing image data is processed by processing Euclidean data,which ignores the non-Euclidean relationship between pixels in hyperspectral remote sensing image.The geomorphic situation expressed by Euclidean sampling is often very limited,so the non-Euclidean relationship between pixels is worth exploring.Therefore,this paper studies the classification of hyperspectral remote sensing images by using graph neural network.In this paper,the research status of hyperspectral remote sensing image classification,graph neural network and superpixel segmentation is firstly described,and then the evolution of spectral convolution,graph neural network model and optimization algorithm is introduced.The specific research content of this paper is as follows:(1)Based on the multi-scale dynamic graph convolutional network(MDGCN)model,an adaptive dynamic multi-scale convolutional network(AMDGCN)model based on super pixel segmentation is proposed.This model can adjust the weight parameters of each scale adaptively according to the process of model optimization,so as to make the model move forward in a more optimized direction in the training process.In addition,two updating methods of scale weights are proposed according to the different degrees of freedom.(2)The improved AMDGCN model was verified with two hyperspectral remote sensing data sets,Indian Pines and University of Pavia.Through the model training results,it is found that the convolution part of graph with scale 2 does not contribute to the model training.Compared with the classification results of the models,the classification results of the AMDGCN model are better than those of MDGCN in the Indian pine dataset,but slightly worse than those of MDGCN in the University of Pavia dataset due to other factors.Therefore,this paper demonstrates that the proposed AMDGCN model is reasonable and effective from two aspects of model contribution and classification results.