沙漠与绿洲气象2025，Vol.19Issue(2)：108-115,8.DOI:10.12057/j.issn.1002-0799.2309.02001
基于改进的DeeplabV3+模型的云检测研究
Research on Cloud Detection Based on Improved DeeplabV3+Model
左昕杰 1武越 1何明元 2张杰 2谢佳桐1
作者信息
- 1. 西北核技术研究所,陕西 西安 710024
- 2. 国防科技大学气象海洋学院,湖南 长沙 410073
- 折叠
摘要
Abstract
Considering that traditional cloud detection methods are difficult to apply to high spatial resolution satellite remote sensing images,this paper employs deep learning to carry out pixel-level cloud detection research for SPOT6/7 satellite remote sensing images.Firstly,a practical software named as CloudLabel was developed based on the cloud labeling method of region growing to build accurate and reasonable datasets for model training and testing.Cloud Label integrates multiple morphological processing methods,which could effectively preserve the edge details of cloud areas.Compared with the existing Labelme labeling CloudLabel improves the pixel-level cloud detection accuracy by 3%.Secondly,the involution module was introduced into the original DeeplabV3+model,and combined with the Poly learning rate change strategy,a cloud detection method based on the RedNet-DeeplabV3+model was proposed.Experimental results demonstrated that the proposed method outperformed other deep learning network models such as DeeplabV3+,U-net and DANet,achieving a cloud detection accuracy of over 93%.In addition,to verify the universality of the proposed method,cloud detection was performed on other types of high spatial resolution remote sensing images such as aerial photos and geographic information system data,and the recognition accuracy exceeded 87.2%.关键词
高空间分辨率卫星遥感影像/云检测/深度学习Key words
high spatial resolution satellite remote sensing images/cloud detection/deep learning分类
大气科学引用本文复制引用
左昕杰,武越,何明元,张杰,谢佳桐..基于改进的DeeplabV3+模型的云检测研究[J].沙漠与绿洲气象,2025,19(2):108-115,8.
