多维卷积和跳跃图卷积融合的高光谱图像分类  

Multi-dimensional convolution and hopping graph convolution fusion for hyperspectral image classification

作　　者：李灏兴 黄平平[1,2] 翟涌光 王志国 董亦凡 郭利彪 LI Haoxing;HUANG Pingping;ZHAI Yongguang;WANG Zhiguo;DONG Yifan;GUO Libiao(College of Information Engineering,Inner Mongolia University of Technology,Hohhot 010051,China;Key Laboratory of Radar Technology and Application,Inner Mongolia Autonomous Region,Hohhot O1oo51,China)

机构地区：[1]内蒙古工业大学信息工程学院,呼和浩特010051 [2]内蒙古自治区雷达技术与应用重点实验室,呼和浩特010051

出　　处：《测绘科学》2024年第6期74-85,共12页Science of Surveying and Mapping

基　　金：国家自然科学基金区域创新发展联合基金重点项目(U22A2010);内蒙古自治区自然科学基金资助项目(2023MS04011);内蒙古自治区科技计划项目(2019GG138);内蒙古自治区直属高校基本科研业务费项目(JY20220072,JY20230008);内蒙古自治区高等学校科学技术研究项目(NJZZ23034)。

摘　　要：针对卷积神经网络应用于小样本分类问题时通常难以达到较高的精度,图卷积网络能够聚合无标签邻域节点特征但未能充分挖掘像素级特征的问题,该文提出了一种融合多维卷积和跳跃图卷积网络的高光谱图像分类方法。该算法包括两个网络分支,首先多维度卷积子网络全面获取空-谱像素级特征;其次跳跃图卷积子网络利用不同层次的邻接矩阵提取全局节点间的关联信息,最后使用交叉注意力融合自适应融合两个子网的特征。实验结果表明,该文所提方法充分结合了卷积神经网络和图卷积网络的优势,在两组广泛使用的高光谱数据集中的分类精度均高于其他对比方法,总体分类精度分别达到了95.85%、96.48%。此外,通过改变网络模型参数验证了该文所提方法的有效性和稳定性。该文为从多层次、多角度提取高光谱图像特征信息用于分类提供了方法参考。Aiming at the problems that convolutional neural networks are usually difficult to achieve high accuracy when applied to small-sample classification problems, and that graph convolutional networks are able to aggregate the features of unlabeled neighborhood nodes but fail to fully exploit the pixel-level features, a hyperspectral image classification method that integrates multidimensional convolution and hopping graph convolutional networks was proposed in this paper. The algorithm consists of two network branches, firstly the multidimensional convolutional sub-network comprehensively acquired the spatial-spectral pixel-level features;secondly the hopping graph convolutional sub-network extracted the global inter-node correlation information by using different levels of neighborhood matrices, and finally the features of the two sub-networks were adaptively fused by using cross-attention fusion. Experimental results showed that the proposed method fully combined the advantages of convolutional neural network and graph convolutional network, and the classification accuracies in the two widely used hyperspectral datasets were higher than the other comparative methods, with the overall classification accuracies reaching 95.85% and 96.48%,respectively. In addition, the effectiveness and stability of the proposed method were verified by changing the network model parameters. This paper provided a methodological reference for extracting hyperspectral image feature information for classification from multiple levels and angles.

关 键 词：卷积神经网络 图卷积网络 高光谱图像分类 多头注意力融合 交叉注意力融合 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置]