基于时序数据回归分析的深度学习云检测  

Deep Learning Cloud Detection Based on Regression Analysis of Temporal Data

作　　者：田亚楠 李云岭[1] 孙林[1] 逄淑林 张平[1] Tian Yanan;Li Yunling;Sun Lin;Pang Shulin;Zhang Ping(College of Geodesy and Geomatics,Shandong University of Science and Technology,Qingdao 266590,Shandong,China)

机构地区：[1]山东科技大学测绘与空间信息学院,山东青岛266590

出　　处：《激光与光电子学进展》2024年第4期464-471,共8页Laser & Optoelectronics Progress

基　　金：科学技术部高端外国专家引进计划(G2021025006L)。

摘　　要：云的存在影响卫星数据的使用,因此准确高效的云检测在遥感图像目标识别和参数定量反演中具有重要作用。针对亮地表、薄碎云及云边界等难以识别、不同尺度特征的云检测精度不稳定等问题,对短期时序数据进行线性回归计算,将前后时序数据的表观反射率斜率变化趋势作为输入特征。为充分挖掘不同尺度的信息,采用具有密集跳跃结构和深监督结构的UNet++模型进行云检测研究。与单时相数据集的U-Net、SegNet和UNet++相比,所提方法可以更有效地突出多尺度特征,增加对亮地表、云边缘和薄云信息的敏感度。结果表明,所提方法在云检测方面得到较高的精度,总体精度达98.21%,误检率降低至1.07%,漏检率降低至3.12%。所提方法能有效降低裸地、道路、建筑物、冰雪等亮地表对云识别的干扰,提升了对薄云的识别精度,且适用于不同下垫面的遥感影像。The transmission of satellite data is adversely affected by cloud cover;therefore,precise cloud detection plays an important role in recognizing remote sensing image targets and quantitatively inverting parameters.This study addresses the challenges of accurately identifying bright surfaces,thin clouds,broken clouds,and cloud boundaries and the stability of cloud detection accuracy across different scale features.We calculate linear regression on short-term time series datasets,using the slope-change trend of apparent reflectance of front and back time series datasets as the input.To fully leverage information from different scales,we employ the UNet++model for cloud detection,which boasts a unique dense skip structure and deep supervision structure.Compared with U-Net,SegNet,and UNet++of the single-temporal dataset,our proposed method can effectively highlight multiscale features and increase the sensitivity for bright surfaces,cloud-boundary contour,and thin-cloud information.Our results demonstrate that the proposed method achieves a high accuracy of 98.21%in cloud detection,and the false detection and missing detection rates are reduced to 1.07%and 3.12%,respectively.Furthermore,our method effectively reduces the interference of bright surfaces on cloud identification,such as barren lands,roads,buildings,ice,and snow,while improving thin-cloud identification accuracy.Therefore,our proposed method is suitable for remote sensing images of different underlying surfaces.

关 键 词：云检测 深度学习 时序数据 线性回归分析 语义分割 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置] P237[自动化与计算机技术—控制科学与工程] P407[天文地球—摄影测量与遥感]