基于局部谱图神经网络的高光谱遥感图像特征可分离性增强及地物分类方法  

Hyperspectral Remote Sensing Image Classification with Feature Separability Enhancement-based Localized Spectral Graph Neural Networks

作　　者：蒲生亮 王济楠 PU Shengliang;WANG Ji'nan(School of Surveying and Geoinformation Engineering,East China University of Technology,Nanchang 330013,China;Key Laboratory of Mine Environmental Monitoring and Improving around Poyang Lake,Ministry of Natural Resources,East China University of Technology,Nanchang 330013,China;Jiangxi Key Laboratory of Watershed Ecological Process and Information,East China University of Technology,Nanchang 330013,China;Nanchang Key Laboratory of Landscape Process and Territorial Spatial Ecological Restoration,East China University of Technology,Nanchang 330013,China;Jiangxi Province Engineering Research Center of Surveying,Mapping and Geographic Information,Nanchang 330013,China)

机构地区：[1]东华理工大学测绘与空间信息工程学院,江西南昌330013 [2]东华理工大学自然资源部环鄱阳湖区域矿山环境监测与治理重点实验室,江西南昌330013 [3]东华理工大学江西省流域生态过程与信息重点实验室,江西南昌330013 [4]东华理工大学南昌市景观过程与国土空间生态修复重点实验室,江西南昌330013 [5]江西省测绘地理信息工程技术研究中心,江西南昌330013

出　　处：《遥感技术与应用》2024年第6期1452-1465,共14页Remote Sensing Technology and Application

基　　金：东华理工大学科研基金项目(DHBK2019192);自然资源部环鄱阳湖区域矿山环境监测与治理重点实验室开放基金项目(MEMI-2021-2022-26);国家自然科学基金项目(42461054)。

摘　　要：采用图几何深度学习方法,具有建模长远程地物特征空间拓扑关系和刻画多地类边界的优势。现有研究多采用主成分分析(Principal Component Analysis,PCA)实现高光谱数据快速降维处理,但是存在特征可分离性较差的问题,使得算法分类结果精度无法继续提升。本研究提出了一种结合图形处理单元(Graphics Processing Unit,GPU)加速的t-分布式随机邻域嵌入(t-distributed Stochastic Neighbor Embedding,t-SNE)流形学习和局部谱域图滤波的高光谱遥感图像分类算法。另一方面,考虑到图注意力网络(Graph Attention Network,GAT)通过利用隐藏的自我注意力层来解决以往基于图卷积或与其近似框架的已知缺点,擅长高效地处理图结构式高光谱数据,继而提出了一种结合局部谱域图滤波和GAT网络的新方法来进行高光谱遥感图像分类。通过在云计算平台Microsoft Planet上使用真实的高光谱遥感数据集进行实验表明,通过高光谱遥感图像特征可分离性增强实现遥感地物分类,不仅能获得较好的高光谱遥感图像地物分类性能,而且说明将空间和光谱信息相结合对高光谱图像分类的重要性。Graph geometrical deep learning has the advantages of modeling topological relationships of long-range ground objects,and describing the boundary of multiple land classes.Existing studies use Principal Component Analysis(PCA)to achieve effective dimensionality reduction of hyperspectral images,but most of them have poor feature separability,which makes the classification performance unable to be further improved.Therefore,the novel hyperspectral remote sensing image classification algorithm based on Graphics Processing Unit(GPU)accelerated t-distributed Stochastic Neighbor Embedding(t-SNE)manifold learning and localized spectral graph filtering was proposed in this study.On the other hand,considering Graph Attention Network(GAT)solves the known shortcomings of previous Graph Convolution Network(GCN)or its approximations by using the hidden self-attention layer,especially since it is good at efficiently processing graph-structured hyperspectral data.Then,the second novel method combining localized spectral graph convolution filtering and GAT network is presented to classify hyperspectral images.Experiments with real hyperspectral datasets on the Microsoft Planet platform show that the proposed methods not only provide new insights into promising hyperspectral image classification performance,but also demonstrate the importance of combining spatial and spectral information for hyperspectral remote sensing image classification.

关 键 词：高光谱图像分类 流形学习 图深度学习 图卷积网络 图注意力网络 

分 类 号：P237[天文地球—摄影测量与遥感]