基于多尺度注意力融合网络的遥感影像场景分类  

Multi-scale attention fusion network for remote sensing image scene classification

作　　者：刘小华 LIU Xiaohua(Fujian Surveying and Mapping Institute,Fuzhou 350003,China)

机构地区：[1]福建省测绘院,福州350003

出　　处：《测绘科学》2024年第12期28-42,共15页Science of Surveying and Mapping

基　　金：福建省自然科学基金项目(2021J01627,2023J01417)。

摘　　要：针对高分辨率遥感影像尺度差异大、空间结构复杂、类内外变异性显著等现象给场景分类带来的技术难题,该文提出了一种基于多尺度注意力融合的深度学习网络。首先,通过ResNet-50提取多尺度特征;然后,利用通道注意力增强类别特征,并通过空间注意力增强目标区域特征;接着,采用串并联集成产生关键语义信息;最后,通过全局跨尺度特征融合模块,增强网络的多尺度特征表达能力,并将特征输入Softmax层实现分类。在公开集AID、NWPU以及福建本地数据集FJ-RSIDS上,本方法分别实现了97.58%、95.03%和98.77%的总体分类精度,以及±0.08%、±0.10%和±0.12%的低分类标准差。相比现有基于卷积网络特征提取的方法,该方法在性能和稳定性上有显著提升,为高分辨率遥感场景分类研究提供了参考。This study addresses challenges such as significant scale variations,complex spatial structures,and high intra-class and inter-class variability in high-resolution remote sensing images to achieve precise and stable scene classification.We propose a Multi-Scale Attention Fusion Network(MSAFNet).First,ResNet-50 is employed for multi-scale feature extraction.Channel attention enhances category features,while spatial attention focuses on target region features.These attention mechanisms are integrated in parallel and serial manners to generate critical semantic information for scene classification.A global cross-scale feature fusion module is introduced to enhance multi-scale feature representation,with the features fed into a Softmax layer for classification.Our method achieved classification accuracies of 97.58%,95.03%,and 98.77%,with low standard deviations of±0.08%,±0.10%,and±0.12%on the AID,NWPU,and FJ-RSIDS datasets,respectively.Compared to existing convolutional network-based methods,our approach significantly improves performance and stability,offering a novel methodology for high-resolution remote sensing scene classification using multi-scale features.

关 键 词：高分辨率遥感 场景分类 深度学习 注意力模块 特征融合 

分 类 号：P237[天文地球—摄影测量与遥感]