利用弱监督预训练CNN模型的遥感影像云检测方法  

Cloud Detection Method for Remote Sensing Image Based on Weakly Supervised Pre-training CNN Model

作　　者：胡文超 解志刚 叶周园 刘亚萍 HU Wenchao;XIE Zhigang;YE Zhouyuan;LIU Yaping(Troops of PLA 32022,Wuhan 430074,China;School of Remote Sensing Information Engineering,Wuhan University,Wuhan 430079,China;CCCC Second Highway Consultants Co.,Ltd.,Wuhan 430056,China)

机构地区：[1]32022部队,湖北武汉430074 [2]武汉大学遥感信息工程学院,湖北武汉430079 [3]中交第二公路勘察设计研究院有限公司,湖北武汉430056

出　　处：《地理空间信息》2024年第12期25-29,共5页Geospatial Information

摘　　要：近年来,深度学习在遥感影像云检测方面得到广泛应用,但该方法依赖大规模高质量的人工标注样本。因此,借鉴弱监督学习和迁移学习思想,提出了一种基于弱监督预训练CNN模型的云检测方法。首先采用快速云检测方法生成大规模弱监督样本,将其代替人工标注样本对CNN模型进行预训练;然后利用少量人工标注样本对所得预训练模型进行精调;最后利用训练所得模型对影像进行云检测。实验结果表明,该方法总体精度达到96.00%,F1分数为87.70%,在维持较好检测精度的前提下显著降低了人工标注成本,在工程实践中具有较好的应用前景。In recent years,deep learning has been widely used in cloud detection of remote sensing images,but it relies on large-scale high-quality manually labeled samples.Based on weakly supervised learning and transfer learning,we proposed a cloud detection method based on weakly supervised pre-training CNN model in this paper.Firstly,we used a fast detection method to generate a large-scale weakly supervised sample,which was used to pre-train the CNN model instead of manually labeled samples.Then,we used a small number of manually labeled samples to fine-tune the pre-training model.Finally,we used the obtained model to conduct cloud detection of images.The experimental results show that the overall accuracy of this method reaches 96.00%,with F1-score of 87.70%.On the premise of maintaining ideal detection accuracy,it significantly reduces manual labeling costs,presenting good application prospects in industrial practice.

关 键 词：云检测 弱监督 CNN 遥感影像 迁移学习 

分 类 号：P237[天文地球—摄影测量与遥感]