高级检索

留言板

尊敬的读者、作者、审稿人, 关于本刊的投稿、审稿、编辑和出版的任何问题, 您可以本页添加留言。我们将尽快给您答复。谢谢您的支持!

姓名
邮箱
手机号码
标题
留言内容
验证码

基于多尺度特征融合网络的云和云阴影检测试验

杨昌军 张秀再 张晨 冯绚 刘瑞霞

杨昌军, 张秀再, 张晨, 等. 2021. 基于多尺度特征融合网络的云和云阴影检测试验[J]. 大气科学, 45(6): 1187−1195 doi: 10.3878/j.issn.1006-9895.2105.20214
引用本文: 杨昌军, 张秀再, 张晨, 等. 2021. 基于多尺度特征融合网络的云和云阴影检测试验[J]. 大气科学, 45(6): 1187−1195 doi: 10.3878/j.issn.1006-9895.2105.20214
YANG Changjun, ZHANG Xiuzai, ZHANG Chen, et al. 2021. Cloud and Cloud Shadow Detection Tests Based on Multiscale Feature Fusion Network [J]. Chinese Journal of Atmospheric Sciences (in Chinese), 45(6): 1187−1195 doi: 10.3878/j.issn.1006-9895.2105.20214
Citation: YANG Changjun, ZHANG Xiuzai, ZHANG Chen, et al. 2021. Cloud and Cloud Shadow Detection Tests Based on Multiscale Feature Fusion Network [J]. Chinese Journal of Atmospheric Sciences (in Chinese), 45(6): 1187−1195 doi: 10.3878/j.issn.1006-9895.2105.20214

基于多尺度特征融合网络的云和云阴影检测试验

doi: 10.3878/j.issn.1006-9895.2105.20214
基金项目: 第二次青藏高原综合科学考察研究项目2019QZKK0105,国家自然科学青年基金项目11504176、61601230、41905033,江苏省自然科学青年基金项目BK20141004,江苏省高校自然科学研究重大项目13KJA510001
详细信息
    作者简介:

    杨昌军,男,1971年出生,副研究员,主要从事卫星大气遥感方面的研究。E-mail: yangcj@cma.gov.cn

    通讯作者:

    张秀再,E-mail: zxzhering@163.com;张晨,E-mail: chenzhang_chn@163.com

  • 中图分类号: P751

Cloud and Cloud Shadow Detection Tests Based on Multiscale Feature Fusion Network

Funds: The Second Tibetan Plateau Scientific Expedition and Research (STEP) Program (Grant 2019QZKK0105), Youth Program of National Natural Science Foundation of China (Grants 11504176, 61601230, 41905033), Youth Program of National Natural Science Foundation of Jiangsu Province (Grant BK20141004), Key University Science Research Project of Jiangsu Province (Grant 13KJA510001)
  • 摘要: 基于深度学习的高分辨率光学影像云检测过程中,云和云阴影及其边缘细节丢失较为严重,主要原因在于不同尺度空间语义信息特征融合存在不足。针对该问题,本文构建一种基于深度学习的多尺度特征融合网络(Multi-scale Feature Fusion Network, MFFN)的云和云阴影检测方法,该算法结合防止网络退化的残差神经网络模块(Res.block)、扩大网络感受野的多尺度卷积模块(MCM)和提取并融合不同尺度信息的多尺度特征模块(MFM)。试验表明,本算法能提取丰富的空间信息与语义信息,可取得较为精细的云与云阴影掩模,具有较高检测精度,其中云检测准确率达0.9796,云阴影检测准确率达0.8307。同时,该工作可为深度学习技术应用于业务云检测提供理论支持及技术储备。
  • 图  1  多尺度特征融合分割网络的残差模块示意图

    Figure  1.  Diagram of the Res.block (residual block) module in the MFFN (multiscale feature fusion network). Conv. 1*1 and Conv. 3*3 denote a 1×1 convolution and 3×3 convolution, respectively, and leaky_ReLU is the activation function

    图  2  多尺度特征融合分割网络的多尺度卷积模块示意图

    Figure  2.  Diagram of the MCM (multiscale convolution module) in the MFFN. Conv. 5*5 denotes a 5×5 convolution

    图  3  多尺度特征融合分割网络的多尺度特征模块示意图

    Figure  3.  Diagram of the MFM (Multiscale feature module) in the MFFN. Up denotes upsample operation, and Concat denotes feature concatenate

    图  4  多尺度特征融合分割网络示意图

    Figure  4.  Diagram of the MFFN. Pooling denotes reducing the size of the feature map

    图  5  Landsat8 SPARCS遥感卫星影像数据的增强操作:(a)RGB三通道合成、(b)垂直旋转、(c)水平旋转、(d)水平并垂直旋转

    Figure  5.  Data augmentation operations of the Landsat8 SPARCS remote sensing satellite image experimental data: (a) RGB tri-channel composite image, (b) vertical flip, (c) horizontal flip, (d) horizontal and vertical flip

    图  6  Landsat8 SPARCS遥感卫星影像数据示例:(a)RGB三通道合成、(b)真值。图b中白色、灰色、黑色分别表示云、云阴影、晴空

    Figure  6.  Examples of the Landsat8 SPARCS remote sensing satellite image data: (a) RGB tri-channel composite image, (b) the truth. In Fig. b, white, gray, and black denote cloud, cloud shadow, and clear sky, respectively

    图  7  云和云阴影检测视觉对比(碎云):(a)RGB三通道合成;(b)K-means检测结果;(c)Resunet检测结果;(d)多尺度特征融合网络检测结果;(e)云和云阴影真实结果

    Figure  7.  Visual comparison of cloud and cloud shadow detection (the small piece of cloud): (a) RGB tri-channel composite image; (b) K-means detection result; (c) resunet detection result; (d) MFFN detection result; (e) cloud and cloud shadow truth result

    图  8  云和云阴影检测视觉对比(冰/雪场景):(a)RGB三通道合成;(b)K-means检测结果;(c)Resunet检测结果;(d)多尺度特征融合网络检测结果;(e)云和云阴影真实结果

    Figure  8.  Visual comparison of cloud and cloud shadow detection (ice/snow scene): (a) RGB tri-channel composite image; (b) K-means detection result; (c) resunet detection result; (d) MFFN detection result; (e) cloud and cloud shadow truth result

    图  9  云阴影检测视觉对比:(a)RGB三通道合成;(b)K-means检测结果;(c)Resunet检测结果;(d)多尺度特征融合网络检测结果;(e)云阴影真实结果

    Figure  9.  Visual comparison of cloud shadow detection: (a) RGB tri-channel composite image; (b) K-means detection result; (c) resunet detection result; (d) MFFN detection result; (e) cloud shadow truth result

    图  10  MFFN及Resunet方法在不同迭代次数时,总体精度(Overall Accueacy)对比曲线

    Figure  10.  Comparison curves of overall accuracy of MFFN and Resunet method at different epochs

    表  1  云检测定量比较

    Table  1.   Quantitative comparison of cloud detection

    方法平均交并比MIoU精确率PPre召回率RRec准确率AAcc调和均值F1score
    K-means0.85360.83820.90680.8458
    Resunet0.82530.93690.91810.96740.9299
    MFFN0.87850.93510.93630.97960.9357
    下载: 导出CSV

    表  2  云阴影检测定量比较

    Table  2.   Quantitative comparison of cloud shadow detection

    方法平均交并比MIoU精确率PPre召回率RRec准确率AAcc调和均值F1score
    K-means0.62910.84330.78020.7206
    Resunet0.80950.78630.94050.81200.8565
    MFFN0.85220.81030.97360.83070.8845
    下载: 导出CSV
  • [1] Angel Y, Houborg R, McCabe M F. 2016. Removal of clouds, dust and shadow pixels from hyperspectral imagery using a non-separable and stationary spatio-temporal covariance model [C]//Proceedings of SPIE 9998, Remote Sensing for Agriculture, Ecosystems, and Hydrology XVIII. Edinburgh: SPIE, 9998: 999824. doi: 10.1117/12.2241518
    [2] Braaten J D, Cohen W B, Yang Z Q. 2015. Automated cloud and cloud shadow identification in Landsat MSS imagery for temperate ecosystems [J]. Remote Sens. Environ., 169: 128−138. doi: 10.1016/j.rse.2015.08.006
    [3] Chai D F, Newsam S, Zhang H K, et al. 2019. Cloud and cloud shadow detection in Landsat imagery based on deep convolutional neural networks [J]. Remote Sens. Environ., 225: 307−316. doi: 10.1016/j.rse.2019.03.007
    [4] Cihlar J, Howarth J. 1994. Detection and removal of cloud contamination from AVHRR images [J]. IEEE Trans. Geosci. Remote Sens., 32(3): 583−589. doi: 10.1109/36.297976
    [5] Fisher A. 2014. Cloud and cloud–shadow detection in SPOT5 HRG imagery with automated morphological feature extraction [J]. Remote Sens, 6(1): 776−800. doi: 10.3390/rs6010776
    [6] Goodwin N R, Collett L J, Denham R J, et al. 2013. Cloud and cloud shadow screening across Queensland, Australia: An automated method for Landsat TM/ETM+ time series [J]. Remote Sens. Environ., 134: 50−65. doi: 10.1016/j.rse.2013.02.019
    [7] He J, Zhang L, Wang Q, et al. 2009. Using diffusion geometric coordinates for hyperspectral imagery representation [J]. IEEE Geosci. Remote Sens. Lett., 6(4): 767−771. doi: 10.1109/LGRS.2009.2025058
    [8] Hu X Y, Wang Y, Shan J. 2015. Automatic recognition of cloud images by using visual saliency features [J]. IEEE Geosci. Remote Sens. Lett., 12(8): 1760−1764. doi: 10.1109/LGRS.2015.2424531
    [9] Hughes M J, Hayes D J. 2014. Automated detection of cloud and cloud shadow in single-date Landsat imagery using neural networks and spatial post-processing [J]. Remote Sens., 6(6): 4907−4926. doi: 10.3390/rs6064907
    [10] Irish R R, Barker J L, Goward S N, et al. 2006. Characterization of the Landsat-7 ETM+ automated cloud-cover assessment (ACCA) algorithm [J]. Photogramm. Eng. Remote Sens., 72(10): 1179−1188. doi: 10.14358/PERS.72.10.1179
    [11] Li P F, Dong L M, Xiao H C, et al. 2015. A cloud image detection method based on SVM vector machine [J]. Neurocomputing, 169: 34−42. doi: 10.1016/j.neucom.2014.09.102
    [12] 梁舒. 2018. 基于残差学习U型卷积神经网络的乳腺超声图像肿瘤分割研究 [D]. 华南理工大学硕士学位论文. Liang S. 2018. Research on breast ultrasound image segmentaion based on residual U-shaped convolution neural network [D]. M. S. thesis (in Chinese), South China University of Technology.
    [13] Liu W, Anguelov D, Erhan D, et al. 2016. SSD: Single shot multibox detector [C]//Proceedings of the 14th European Conference on Computer Vision. Amsterdam: Springer, 21–37. doi: 10.1007/978-3-319-46448-0_2
    [14] Mohajerani S, Krammer T A, Saeedi P. 2018. A cloud detection algorithm for remote sensing images using fully convolutional neural networks [C]//Proceedings of the 2018 IEEE 20th International Workshop on Multimedia Signal Processing. Vancouver: IEEE. doi: 10.1109/MMSP.2018.8547095
    [15] Oreopoulos L, Wilson M J, Várnai T. 2011. Implementation on Landsat data of a simple cloud-mask algorithm developed for MODIS land bands [J]. IEEE Geosci. Remote Sens. Lett., 8(4): 597−601. doi: 10.1109/LGRS.2010.2095409
    [16] Ozkan S, Efendioglu M, Demirpolat C. 2018. Cloud detection from RGB color remote sensing images with deep pyramid networks [C]//Proceedings of 2018 IEEE International Geoscience and Remote Sensing Symposium. Valencia: IEEE, 6939–6942. doi: 10.1109/IGARSS.2018.8519570
    [17] Shao Z F, Deng J, Wang L, et al. 2017. Fuzzy autoencode based cloud detection for remote sensing imagery [J]. Remote Sens., 9(4): 311. doi: 10.3390/rs9040311
    [18] Shi M Y, Xie F Y, Zi Y, et al. 2016. Cloud detection of remote sensing images by deep learning [C]//Proceedings of 2016 IEEE International Geoscience and Remote Sensing Symposium. Beijing: IEEE, 701–704. doi: 10.1109/IGARSS.2016.7729176
    [19] Tan K, Zhang Y J, Tong X. 2016. Cloud extraction from Chinese high resolution satellite imagery by probabilistic latent semantic analysis and object-based machine learning [J]. Remote Sens., 8(11): 963. doi: 10.3390/rs8110963
    [20] van Zyl Marais I, Du Preez J A, Steyn W H. 2011. An optimal image transform for threshold-based cloud detection using heteroscedastic discriminant analysis [J]. Int. J. Remote Sens., 32(6): 1713−1729. doi: 10.1080/01431161003621619
    [21] 王权, 孙林, 韦晶, 等. 2018. 动态阈值云检测算法改进及在高分辨率卫星上的应用 [J]. 光学学报, 38(10): 1028002. doi: 10.3788/AOS201838.1028002

    Wang Q, Sun L, Wei J, et al. 2018. Improvement of universal dynamic threshold cloud detection algorithm and its application in high resolution satellite [J]. Acta Opt. Sinica, 38(10): 1028002. doi: 10.3788/AOS201838.1028002
    [22] Xu K, Guan K Y, Peng J, et al. 2019. DeepMask: An algorithm for cloud and cloud shadow detection in optical satellite remote sensing images using deep residual network. arXiv, 1911. 03607. Available athttps://arxiv.org/abs/1911.03607 [2020-08-12]
    [23] Xu L N, Niu R Q, Fang S H, et al. 2013. Cloud detection based on decision tree over Tibetan Plateau with MODIS data [C]//Proceedings of SPIE 8921, Remote Sensing Image Processing, Geographic Information Systems, and Other Applications. Wuhan: SPIE, 89210G. doi: 10.1117/12.2030399
    [24] 杨昌军, 许健民, 赵凤生. 2008. 时间序列在FY2C云检测中的应用 [J]. 大气与环境光学学报, 3(5): 377−391. doi: 10.3969/j.issn.1673-6141.2008.05.010

    Yang C J, Xu J M, Zhao F S. 2008. Application of time series in FY2C cloud detection [J]. J. Atmos. Environ. Opt., 3(5): 377−391. doi: 10.3969/j.issn.1673-6141.2008.05.010
    [25] Wu X, Shi Z W. 2018. Utilizing multilevel features for cloud detection on satellite imagery [J]. Remote Sens., 10(11): 1853. doi: 10.3390/rs10111853
    [26] Zhao H S, Shi J P, Qi X J, et al. 2017. Pyramid scene parsing network [C]//Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition. Honolulu: IEEE, 6230–6239. doi: 10.1109/CVPR.2017.660
    [27] 张建萍, 刘希玉. 2007. 基于聚类分析的K-means算法研究及应用 [J]. 计算机应用研究, 24(5): 166−168. doi: 10.3969/j.issn.1001-3695.2007.05.051

    Zhang J P, Liu X Y. 2007. Application in cluster’s analysis is analyzed in children development period [J]. Appl. Res. Comput., 24(5): 166−168. doi: 10.3969/j.issn.1001-3695.2007.05.051
    [28] Zhu X L, Helmer E H. 2018. An automatic method for screening clouds and cloud shadows in optical satellite image time series in cloudy regions [J]. Remote Sens. Environ., 214: 135−153. doi: 10.1016/j.rse.2018.05.024
  • 加载中
图(10) / 表(2)
计量
  • 文章访问数:  96
  • HTML全文浏览量:  21
  • PDF下载量:  36
  • 被引次数: 0
出版历程
  • 收稿日期:  2020-10-16
  • 录用日期:  2021-05-28
  • 网络出版日期:  2021-05-31
  • 刊出日期:  2021-11-25

目录

    /

    返回文章
    返回