测绘通报 ›› 2020, Vol. 0 ›› Issue (3): 17-20,34.doi: 10.13474/j.cnki.11-2246.2020.0070

• 学术研究 • 上一篇    下一篇

基于改进U-Net网络的遥感图像云检测

张永宏1, 蔡朋艳1, 陶润喆1, 王剑庚2, 田伟3   

  1. 1. 南京信息工程大学自动化学院, 江苏 南京 210044;
    2. 南京信息工程大学大气物理学院, 江苏 南京 210044;
    3. 南京信息工程大学计算机与软件学院, 江苏 南京 210044
  • 收稿日期:2019-08-21 修回日期:2019-10-18 发布日期:2020-04-09
  • 通讯作者: 蔡朋艳。E-mail:2356852131@qq.com E-mail:2356852131@qq.com
  • 作者简介:张永宏(1974-),男,博士,教授,主要从事遥感图像处理分析研究。E-mail:zyh@nuist.edu.cn
  • 基金资助:
    国家自然科学基金(41661144039;41875027)

Cloud detection for remote sensing images using improved U-Net

ZHANG Yonghong1, CAI Pengyan1, TAO Runzhe1, WANG Jiangeng2, TIAN Wei3   

  1. 1. School of Automation, Nanjing University of Information&Technology, Nanjing 210044, China;
    2. School of Atmospheric Physics, Nanjing University of Information&Technology, Nanjing 210044, China;
    3. School of Computer and Software, Nanjing University of Information&Technology, Nanjing 210044, China
  • Received:2019-08-21 Revised:2019-10-18 Published:2020-04-09

摘要: 为了解决U-Net模型应用于云检测时对碎云和薄云存在漏检的问题,本文提出了一种改进的U-Net网络模型,并应用于FY-4A数据进行云检测。首先,利用国家气象卫星中心提供的云检测产品生成二分类云标签;其次,将U-Net模型的编码器与残差模块相结合,使得网络参数共享,并避免深层网络的退化问题;最后,在解码器中融入密集连接模块,将浅层特征与深层特征进行连接,便于获取新的特征,并提高特征使用率。试验结果表明,模型在测试集上的IOU值和Dice系数分别为91.5%和95.2%,可以很好地检测出薄云及大量碎云,效果明显优于U-Net模型。

关键词: 云检测, U-Net, 残差模块, 密集连接模块, FY-4A

Abstract: An improved U-Net model is proposed to solve the problem that missing detection of fragmentary clouds and thin clouds when U-Net is applied to detect clouds, and applied to cloud detection of FY-4A data. Firstly, the cloud inspection product of the National Meteorological Satellite Center is used to generate binary cloud label. Secondly, the encoder of U-Net is combined with residual block to share parameters and avoid degradation of deep network. Finally, the dense block is integrated into the decoder to connect the shallow features with the deep features, which is conducive to acquiring new features and improving the utilization rate of features. The experimental results show that the IOU and Dice coefficients of the model on the test set are 91.5% and 95.2% respectively, which can detect thin clouds and a large number of broken clouds well, and the effect is obviously better than that of the U-Net model.

Key words: cloud detection, U-Net, residual block, dense block, FY-4A

中图分类号: