葛小三, 郑猛猛.基于SER-GNN的小样本遥感影像分类研究[J].河南理工大学学报（自然科学版）,2025,44(5):144-151.

GE X S, ZHENG M M.Study on classification of few-shot remote sensing images based on SER-GNN[J].Journal of Henan Polytechnic University(Natural Science) ,2025,44(5):144-151.

基于SER-GNN的小样本遥感影像分类研究

葛小三1,2, 郑猛猛1,2

1.河南理工大学 测绘与国土信息工程学院，河南 焦作  454000；2.河南理工大学 自然资源部矿山时空信息与生态修复重点实验室，河南 焦作  454000

摘要: 目的 为解决基于度量学习的遥感影像分类中小样本学习特征空间图像特征分布不明显问题，提出一种适用于小样本模型的遥感影像分类模型SER-GNN（SENet attention residual neural network and graph neural networks）。  方法 该模型首先通过SER-GNN卷积层（融合基础网络ResNet-12和SENet组成）进行遥感影像图像特征提取，增强模型对关键区域的关注能力；然后将图像信息和类别标签编码嵌入到SER-GNN模型的GNN层；最后以消息传递推理算法的模式计算影像类别之间的隐含关系，构建邻接网络并完成分类任务。  结果 结果表明，该模型在UC Merced Land-Use数据集、AID遥感数据集、NWPU-RESISC45数据集上，在5-way 1-shot中，精度分别提高1.35%，2.15%，1.3%；在5-way 5-shot中精度分别提高2.15%，5.65%，3.85%。此外，通过迁移学习策略，在NWPU-RESISC45上训练的模型在AID和UC Merced Land-Use数据集上展现出更优的泛化性能。  结论 综上，本文提出的SER-GNN 模型有效融合卷积神经网络与图神经网络的结构优势，在遥感影像小样本分类任务中表现出更高的准确率的同时，在模型迁移上取得了更强的迁移适应能力。该模型在新的学习环境中获得了更好的适应性，为遥感影像智能分类提供了具有潜力的技术路径与方法参考。

关键词:影像分类;小样本学习;ResNet-12;图神经网络;节点嵌入

DOI:10.16186/j.cnki.1673-9787.2023060037

基金项目:国家自然科学基金资助项目（41572341）；河南省自然科学基金资助项目（222300420450）；自然资源部矿山时空信息与生态修复重点实验室开放基金资助项目（KLM202319）

收稿日期:2023/06/16

修回日期:2023/11/02

出版日期:2025/07/23

Study on classification of few-shot remote sensing images based on SER-GNN

Ge Xiaosan1,2, Zheng Mengmeng1,2

1.School of Surveying and Mapping and Land Information Engineering， Henan Polytechnic University， Jiaozuo  454000， Henan， China；2.Key Laboratory of Mine Spatio-Temporal Information and Ecological Restoration， MNR， Henan Polytechnic University， Jiaozuo  454000， Henan， China

Abstract: Objectives However， metric learning-based approaches often suffer from unclear feature distribution in the feature space under few-shot settings.To address this issue， a few-shot remote sensing image classification model named SER-GNN （SENet attention residual neural network and graph neural networks） was proposed.  Methods The model utilized attention mechanisms to enhance feature extraction capabilities through residual networks. Remote sensing image data was first processed by the SER-GNN convolutional layer （composed of ResNet-12 and SENet） to extract image features and enhance attention to key regions. Then， the image information and category label encodings were embedded into the GNN layer of the SER-GNN model. Finally， the implicit relationships between image categories were inferred using a message passing algorithm to construct an adjacency network and complete the classification task.  Results Experiments showed that the model improved classification accuracy by 1.35%， 2.15%， and 1.3% on the UC Merced Land-Use， AID， and NWPU-RESISC45 datasets under the 5-way 1-shot setting， respectively； and by 2.15%， 5.65%， and 3.85% under the 5-way 5-shot setting， respectively. Moreover， through a transfer learning strategy， the model trained on the NWPU-RESISC45 dataset demonstrated superior generalization performance when applied to the AID and UC Merced Land-Use datasets. Conclusions The proposed SER-GNN model effectively combined the structural advantages of convolutional and graph neural networks. It achieved higher classification accuracy in few-shot remote sensing tasks and demonstrated better adaptability in transfer learning. The results suggested that the model could adapt well to new learning environments and provided a promising technical path and methodological reference for intelligent classification of remote sensing images under few-shot conditions.

Key words:image classification;few-shot learning;ResNet-12;graph neural network;node embedding