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

基于软硬协作决策的半监督珊瑚礁底质分类方法

于俊 陈辉 朱大明 程亮 段志鑫 庄启智 楚森森 杨伟 杜思雨

于俊,陈辉,朱大明,等. 基于软硬协作决策的半监督珊瑚礁底质分类方法[J]. 海洋学报,2023,45(4):154–164 doi: 10.12284/hyxb2023049
引用本文: 于俊,陈辉,朱大明,等. 基于软硬协作决策的半监督珊瑚礁底质分类方法[J]. 海洋学报,2023,45(4):154–164 doi: 10.12284/hyxb2023049
Yu Jun,Chen Hui,Zhu Daming, et al. A semi-supervised coral reef substrate classification method based on soft and hard collaborative decision making[J]. Haiyang Xuebao,2023, 45(4):154–164 doi: 10.12284/hyxb2023049
Citation: Yu Jun,Chen Hui,Zhu Daming, et al. A semi-supervised coral reef substrate classification method based on soft and hard collaborative decision making[J]. Haiyang Xuebao,2023, 45(4):154–164 doi: 10.12284/hyxb2023049

基于软硬协作决策的半监督珊瑚礁底质分类方法

doi: 10.12284/hyxb2023049
基金项目: 国家自然科学基金(42001401)。
详细信息
    作者简介:

    于俊(1994-),男,江西省九江市人,研究方向为遥感、地理信息系统。E-mail: 2571445820@qq.com

    通讯作者:

    朱大明(1970-),博士,副教授,研究方向为地理信息系统、空间分析、3S集成。E-mail: 634617255@qq.com

  • 中图分类号: P737.22

A semi-supervised coral reef substrate classification method based on soft and hard collaborative decision making

  • 摘要: 珊瑚礁底质分类对海洋资源开发和海洋生态环境保护起到至关重要的作用。目前,深度学习语义分割方法在遥感图像分类领域应用广泛,但在底质分类方面的研究较少。由于基于全监督深度学习的方法中逐像素标注标签的成本较高,不适用于大规模、高频次的底质分类工作,基于半监督的深度学习方法能够有效利用已标注标签为无标签数据产生伪标签,从而有效降低人工成本,然而现有半监督方法的性能易受伪标签噪声的干扰。针对以上问题,本文提出了一种基于软硬协作决策的半监督底质分类方法。首先,利用多模型联合决策生成高质量的伪标签;然后,提出了一种能够顾及伪标签像素置信度的损失函数来指导模型进行训练;最后,采用软硬协作的决策方式得到精确的底质分类结果。在美属维尔京群岛圣克罗伊岛北部的巴克岛礁和夏威夷群岛的中途岛东南约400 km处的珍珠与爱马仕环礁的浅层底栖生物栖息地地图数据集上评估了本文方法的精度,实验结果表明,本文提出的方法与全监督学习方法精度相当,比主流的语义分割方法精度平均高3.08%,能够有效服务于珊瑚礁底质调查工作。
  • 图  1  数据集影像(a, b)与标签数据(c, d)(1 mi≈1.609 km)

    Fig.  1  Data set image (a, b) with label data (c, d) (1 mi≈1.609 km)

    图  2  基于多模型联合决策的伪标签生成

    Fig.  2  Pseudo-label generation based on joint multi-model decision making

    图  3  3CLoss指导下的软硬协作底质分类

    Fig.  3  3CLoss-guided classification of soft and hard collaborative substrates

    图  4  BIRNM、PHA数据集的参考结果(a, c)和本文方法底质分类结果(b, d)(1 mi≈1.609 km)

    Fig.  4  Reference results of BIRNM, PHA datasets (a, c) and substrate classification results of the method in this paper (b, d) (1 mi≈1.609 km)

    图  5  不同网络在BIRNM数据上的语义分割结果

    Fig.  5  Semantic segmentation results of different networks on BIRNM data

    图  6  不同方法在PHA数据集上的语义分割结果

    Fig.  6  Semantic segmentation results of different networks on PHA dataset

    图  7  不同底质分类方法整体效果图(左列为BIRNM,右列为PAH)(1 mi≈1.609 km)

    Fig.  7  Overall effect of different substrate classification methods (the left column are BIRNM, right column are PAH) (1 mi≈1.609 km)

    表  1  BIRNM数据集底质分类类型描述

    Tab.  1  Description of substrate classification types in the BIRNM dataset

    类别描述卫星水下
    珊瑚连续的、高浮雕式的珊瑚形成,形状各异,包括平行于大陆架边缘的线性珊瑚
    碎屑死去的、不稳定的珊瑚瓦砾,经常被丝状物或其他大型藻类所占据。这种底质经常出现在礁顶,珊瑚礁碎石可以在宽阔的近海沙地上以低密度聚集的方式出现
    岩石从岛屿基岩延伸到海上的固体碳酸盐块的聚集,或从原生床剥离和运输的松散碳酸盐碎片,根据温特沃斯标准,单个巨石的直径在0.25~3 m之间
    粗糙的沉积物,通常存在于海流或波浪能量影响的区域。颗粒大小在
    1/16~256 mm不等
    下载: 导出CSV

    表  2  巴克岛礁数据集半监督实验结果

    Tab.  2  Results of semi-supervised experiments on the Buck Island dataset

    方法背景珊瑚碎屑岩石mIoU
    Lawin0.930 00.853 20.729 80.367 00.740 00.723 3
    SegFormer0.942 80.863 10.689 50.513 10.762 50.754 2
    PanopticDeep0.930 40.851 20.679 00.442 50.713 80.718 3
    本文方法0.939 50.869 10.726 10.539 90.784 70.771 9
    下载: 导出CSV

    表  3  珍珠与爱马仕环礁半监督实验结果

    Tab.  3  Results of semi-supervised experiments on the Pearl and Hermes Atoll

    方法背景珊瑚碎屑mIoU
    Lawin0.874 10.675 50.528 50.452 10.632 5
    SegFormer0.872 50.661 20.523 10.437 10.623 5
    PanopticDeep0.866 60.672 60.497 50.443 10.619 9
    本文方法0.875 30.684 20.557 70.472 10.647 3
    下载: 导出CSV

    表  4  3CLoss对实验精度的影响

    Tab.  4  3CLoss effect on experimental accuracy

    方法Loss背景珊瑚碎屑岩石mIoU
    本文方法 3CLoss 0.939 5 0.869 1 0.726 1 0.539 9 0.784 7 0.771 9
    本文方法 Cross Entropy Loss 0.942 7 0.859 8 0.713 7 0.477 9 0.746 4 0.748 1
    下载: 导出CSV

    表  5  迭代次数对平均交并比的影响

    Tab.  5  Effect of number of iterations on mIoU

    迭代次数背景珊瑚碎屑岩石mIoU
    10.935 30.867 00.729 80.545 00.774 20.770 3
    20.939 50.869 10.726 10.539 90.784 70.771 9
    30.937 80.866 90.733 80.531 80.772 10.768 4
    下载: 导出CSV

    表  6  本文方法与全监督语义分割方法对比结果

    Tab.  6  Comparison results between the method in this paper and the fully supervised semantic segmentation method

    方法监督方法mIoU(BIRNM)mIoU(PHA)
    Lawin全监督0.759 60.633 5
    SegFormer全监督0.770 60.638 1
    PanopticDeep全监督0.743 80.624 7
    本文方法半监督0.771 90.647 3
    下载: 导出CSV
  • [1] 逄岩, 许枫, 刘佳. 基于Gammatone滤波器组时频谱和卷积神经网络的海底底质分类[J]. 应用声学, 2021, 40(4): 510−517. doi: 10.11684/j.issn.1000-310X.2021.04.003

    Pang Yan, Xu Feng, Liu Jia. Seabed sediment classification based on Gammatone filter banks time-frequency spectrum and convolutional neural networks[J]. Journal of Applied Acoustics, 2021, 40(4): 510−517. doi: 10.11684/j.issn.1000-310X.2021.04.003
    [2] Gregr E J, Haggarty D R, Davies S C, et al. Comprehensive marine substrate classification applied to Canada’s Pacific shelf[J]. PLoS One, 2021, 16(10): e0259156. doi: 10.1371/journal.pone.0259156
    [3] Reshitnyk L, Costa M, Robinson C, et al. Evaluation of WorldView-2 and acoustic remote sensing for mapping benthic habitats in temperate coastal Pacific waters[J]. Remote Sensing of Environment, 2014, 153: 7−23. doi: 10.1016/j.rse.2014.07.016
    [4] Wicaksono P, Aryaguna P A. Analyses of inter-class spectral separability and classification accuracy of benthic habitat mapping using multispectral image[J]. Remote Sensing Applications: Society and Environment, 2020, 19: 100335. doi: 10.1016/j.rsase.2020.100335
    [5] Dempster A P, Laird N M, Rubin D B. Maximum likelihood from incomplete data via the EM algorithm[J]. Journal of the Royal Statistical Society: Series B (Methodological), 1977, 39(1): 1−22. doi: 10.1111/j.2517-6161.1977.tb01600.x
    [6] Pillay T, Cawthra H C, Lombard A T. Characterisation of seafloor substrate using advanced processing of multibeam bathymetry, backscatter, and sidescan sonar in Table Bay, South Africa[J]. Marine Geology, 2020, 429: 106332. doi: 10.1016/j.margeo.2020.106332
    [7] Blaschke T. Object based image analysis for remote sensing[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2010, 65(1): 2−16. doi: 10.1016/j.isprsjprs.2009.06.004
    [8] Furey T S, Cristianini N, Duffy N, et al. Support vector machine classification and validation of cancer tissue samples using microarray expression data[J]. Bioinformatics, 2000, 16(10): 906−914. doi: 10.1093/bioinformatics/16.10.906
    [9] Breiman L. Random forests[J]. Machine Learning, 2001, 45(1): 5−32. doi: 10.1023/A:1010933404324
    [10] 万佳馨, 任广波, 马毅. 基于WorldView-2和GF-2遥感影像的赵述岛礁坪底质变化研究[J]. 海洋科学, 2019, 43(10): 43−54.

    Wan Jiaxin, Ren Guangbo, Ma Yi. Study on substrate changes of Zhaoshu reef flat based on WorldView-2 and GF-2 remote sensing images[J]. Marine Sciences, 2019, 43(10): 43−54.
    [11] 逄今朝, 任广波, 施祺, 等. 基于底质类型变化监测的2005−2018年西沙永乐群岛珊瑚礁白化分析[J]. 海洋科学, 2021, 45(6): 92−106.

    Pang Jinzhao, Ren Guangbo, Shi Qi, et al. Analysis of coral reef bleaching in Yongle Islands of Xisha from 2005 to 2018 based on sediment types change monitoring[J]. Marine Sciences, 2021, 45(6): 92−106.
    [12] 董娟, 任广波, 胡亚斌, 等. 基于高分辨率遥感的珊瑚礁地貌单元体系构建和分类方法——以8波段Worldview-2影像为例[J]. 热带海洋学报, 2020, 39(4): 116−129.

    Dong Juan, Ren Guangbo, Hu Yabin, et al. Construction and classification of coral reef geomorphic unit system based on high-resolution remote sensing: using 8-band Worldview-2 Image as an example[J]. Journal of Tropical Oceanography, 2020, 39(4): 116−129.
    [13] 李晓敏, 马毅, 吕喜玺. 南海珊瑚岛礁遥感分类体系和解译标志[J]. 海洋科学, 2021, 45(5): 23−30. doi: 10.11759/hykx20201110003

    Li Xiaomin, Ma Yi, Lü Xixi. Establishing a remote sensing classification system and interpretation marks for the coral islands and reefs in the South China Sea[J]. Marine Sciences, 2021, 45(5): 23−30. doi: 10.11759/hykx20201110003
    [14] Wan Jiaxin, Ma Yi. Multi-scale spectral-spatial remote sensing classification of coral reef habitats using CNN-SVM[J]. Journal of Coastal Research, 2020, 102(S1): 11−20.
    [15] Huang Rongyong, Zhang Huiya, Yu Kefu. Analysis on the live coral cover around Weizhou Island using MODIS data[J]. Sensors, 2019, 19(19): 4309. doi: 10.3390/s19194309
    [16] 谭琨, 王雪, 杜培军. 结合深度学习和半监督学习的遥感影像分类进展[J]. 中国图象图形学报, 2019, 24(11): 1823−1841. doi: 10.11834/jig.190348

    Tan Kun, Wang Xue, Du Peijun. Research progress of the remote sensing classification combining deep learning and semi-supervised learning[J]. Journal of Image and Graphics, 2019, 24(11): 1823−1841. doi: 10.11834/jig.190348
    [17] King A, Bhandarkar S M, Hopkinson B M. A comparison of deep learning methods for semantic segmentation of coral reef survey images[C]//Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. Salt Lake City: IEEE, 2018: 1394−1402.
    [18] Li Jiwei, Knapp D E, Fabina N S, et al. A global coral reef probability map generated using convolutional neural networks[J]. Coral Reefs, 2020, 39(6): 1805−1815. doi: 10.1007/s00338-020-02005-6
    [19] Wang Mengqiu, Hu Chuanmin. Satellite remote sensing of pelagic Sargassum macroalgae: the power of high resolution and deep learning[J]. Remote Sensing of Environment, 2021, 264: 112631. doi: 10.1016/j.rse.2021.112631
    [20] 耿艳磊, 陶超, 沈靖, 等. 高分辨率遥感影像语义分割的半监督全卷积网络法[J]. 测绘学报, 2020, 49(4): 499−508.

    Geng Yanlei, Tao Chao, Shen Jing, et al. High-resolution remote sensing image semantic segmentation based on semi-supervised full convolution network method[J]. Acta Geodaetica et Cartographica Sinica, 2020, 49(4): 499−508.
    [21] 李鑫伟, 李彦胜, 张永军. 弱监督深度语义分割网络的多源遥感影像水体检测[J]. 中国图象图形学报, 2021, 26(12): 3015−3026.

    Li Xinwei, Li Yansheng, Zhang Yongjun. Weakly supervised deep semantic segmentation network for water body extraction based on multi-source remote sensing imagery[J]. Journal of Image and Graphics, 2021, 26(12): 3015−3026.
    [22] Xie Enze, Wang Wenhai, Yu Zhiding, et al. SegFormer: simple and efficient design for semantic segmentation with transformers[J]. Advances in Neural Information Processing Systems, 2021, 34: 12077−12090.
    [23] Cheng Bowen, Collins M D, Zhu Yukun, et al. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. New York: IEEE, 2020: 12475−12485.
    [24] Yan Haotian, Zhang Chuang, Wu Ming. Lawin transformer: improving semantic segmentation transformer with multi-scale representations via large window attention[EB/OL]. (2022–01–05)[2022–08–15]. https://arxiv.org/abs/2201.01615.
  • 加载中
图(7) / 表(6)
计量
  • 文章访问数:  322
  • HTML全文浏览量:  117
  • PDF下载量:  33
  • 被引次数: 0
出版历程
  • 收稿日期:  2022-08-16
  • 修回日期:  2022-10-22
  • 网络出版日期:  2023-03-29
  • 刊出日期:  2023-03-31

目录

    /

    返回文章
    返回