欢迎访问中国科学院大学学报,今天是

中国科学院大学学报 ›› 2022, Vol. 39 ›› Issue (5): 712-720.DOI: 10.7523/j.ucas.2020.0036

• 电子信息与计算机科学 • 上一篇    

勺型网络:用于Landsat遥感图像云检测的新型网络

王树立1,2, 唐海蓉1,2, 计璐艳1   

  1. 1. 中国科学院空天信息创新研究院 中国科学院空间信息处理与应用系统技术重点实验室, 北京 100094;
    2. 中国科学院大学, 北京 100049
  • 收稿日期:2020-04-17 修回日期:2020-07-21 发布日期:2021-06-01
  • 通讯作者: 王树立
  • 基金资助:
    国家自然科学基金(61701477、61805246)和国家重点研发计划(2017YFB0502903)资助

Spoon network: a new network structure for Landsat imagery cloud detection

WANG Shuli1,2, TANG Hairong1,2, JI Luyan1   

  1. 1. CAS Key Laboratory of Technology in Geo-spatial Information Processing and Application System, Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100094, China;
    2. University of Chinese Academy of Sciences, Beijing 100049, China
  • Received:2020-04-17 Revised:2020-07-21 Published:2021-06-01

摘要: 针对目前用于遥感图像云检测的神经网络模型存在光谱信息未能充分利用而导致的细节信息易损失、碎云漏检率大、计算复杂等不足,提出一种新型且轻量的网络,称为勺型网络(spoon-net,S-Net),应用于Landsat遥感图像的云检测。S-Net分为2个阶段,第1阶段,使用1×1的卷积核提取图像光谱特征,避免图像细节被模糊;第2阶段,使用encoder-decoder框架提取图像空间特征,并引入分组卷积,对第1阶段提取的每一层光谱通道单独进行卷积,保持光谱特征并减少模型参数。模型在Landsat8 biome数据训练测试并评估,结果表明模型在内存与时间上具有较大优势,并达到95%的准确率。

关键词: Landsat, 云检测, 神经网络, 光谱特征, 勺型网络

Abstract: In view of the shortcomings of the neural network model for remote sensing image cloud detection, such as the loss of detail information, the high cloud miss detection rate and the complexity of calculation caused by the insufficient utilization of spectral information, this paper proposes a new and lightweight network called spoon net (S-Net), which is applied to the cloud detection of Landsat remote sensing image. S-Net is divided into two stages. In the first stage, the convolution kernel of 1×1 is used to extract image spectral features to avoid image details being blurred; in the second stage, the encoder decoder framework is used to extract image spatial features, and group convolution is introduced to convolute each layer of spectral channels extracted in the first stage separately to maintain spectral features and reduce model parameters. The model is trained and evaluated in Landsat8 biome dataset, and the results show that the model has a great advantage in memory and time, and achieves an accuracy of 95%.

Key words: Landsat, cloud detection, neural network, spectral characteristics, spoon-net

中图分类号: