激光导航与决策融合的小样本无人机航拍图像分类  

Enhancing few-shot UAV image classification with laser navigation-assisted decision fusion

作　　者：谢幸生 张永挺 丁宗宝 江玉欢 刘剑[2] XIE Xingsheng;ZHANG Yongting;DING Zongbao;JIANG Yuhuan;LIU Jian(Zhongshan Power Supply Bureau of Guangdong Power Grid Co.,Ltd.,Zhongshan 528400,China;School of Electrical Engineering and Automation,Wuhan University,Wuhan 430072,China)

机构地区：[1]广东电网有限责任公司中山供电局,广东中山528400 [2]武汉大学电气与自动化学院,湖北武汉430072

出　　处：《测绘通报》2024年第7期173-177,共5页Bulletin of Surveying and Mapping

基　　金：南方电网公司科技项目(GDKJXM20230706)。

摘　　要：近年来,激光导航技术与小样本无人机航拍图像分类,为土地利用调查、城市规划和环境监测等领域提供了精确的空间定位与大量价值信息,显著提升了分类技术的水平。本文提出了一种激光导航与决策融合技术的小样本无人机航拍图像分类方法,旨在提高分类性能与空间定位精度。通过激光导航系统提供的高精度地理位置信息,优化了航拍图像的特征提取过程,采用自监督学习构建辅助任务,通过旋转和翻转技术增强特征提取器的泛化能力。此外,结合两种自监督范式训练得到的特征提取器,通过逻辑回归分类器完成分类任务,设计了一种新型的决策融合模块,以自动调整各决策权重,提高了分类准确性。通过NWPU-RESISC45和UC Merced数据集上进行试验,结果验证了本文方法的有效性和先进性,展现了激光导航技术在提高小样本无人机航拍图像分类中的潜力。In recent years,the integration of laser navigation technology with few-shot UAV aerial image classification has provided unprecedented precise spatial positioning and valuable information for fields such as land use survey,urban planning,and environmental monitoring,significantly enhancing the importance of classification technology applications.This study proposes a method of few-shot UAV aerial image classification that integrates laser navigation and decision fusion techniques,aimed at improving classification performance and spatial positioning accuracy.By utilizing the high-precision geographic location information provided by the laser navigation system,the feature extraction process of aerial images is optimized.The study adopts self-supervised learning to construct auxiliary tasks,enhancing the generalization ability of feature extractors through rotation and flipping techniques.Moreover,combining feature extractors trained with two self-supervised paradigms and utilizing a logistic regression classifier for the classification task.A novel decision fusion module is designed to automatically adjust the weights of each decision,enhancing classification accuracy.Experimental results on the NWPU-RESISC45 and UC Merced datasets validate the effectiveness and advanced nature of the proposed method,demonstrate the potential of laser navigation technology in enhancing few-shot UAV aerial image classification.

关 键 词：小样本 无人机航拍图像分类 决策融合 自监督学习 激光导航 

分 类 号：P237[天文地球—摄影测量与遥感]