基于监督对比学习正则化的高分辨率SAR图像建筑物提取方法

康健 王智睿 祝若鑫 孙显

康健, 王智睿, 祝若鑫, 等. 基于监督对比学习正则化的高分辨率SAR图像建筑物提取方法[J]. 雷达学报, 2022, 11(1): 157–167. doi: 10.12000/JR21124
引用本文: 康健, 王智睿, 祝若鑫, 等. 基于监督对比学习正则化的高分辨率SAR图像建筑物提取方法[J]. 雷达学报, 2022, 11(1): 157–167. doi: 10.12000/JR21124
KANG Jian, WANG Zhirui, ZHU Ruoxin, et al. Supervised contrastive learning regularized high-resolution synthetic aperture radar building footprint generation[J]. Journal of Radars, 2022, 11(1): 157–167. doi: 10.12000/JR21124
Citation: KANG Jian, WANG Zhirui, ZHU Ruoxin, et al. Supervised contrastive learning regularized high-resolution synthetic aperture radar building footprint generation[J]. Journal of Radars, 2022, 11(1): 157–167. doi: 10.12000/JR21124

基于监督对比学习正则化的高分辨率SAR图像建筑物提取方法

DOI: 10.12000/JR21124
基金项目: 国家自然科学基金(62101371, 62076241),江苏省青年基金项目(BK20210707)
详细信息
    作者简介:

    康 健(1991–),男,2019年在慕尼黑工业大学获得博士学位,现任苏州大学电子信息学院副教授,硕士生导师,IEEE会员,雷达学报客座编辑。主要研究方向为遥感图像智能解译

    王智睿(1990–),男,2018年在清华大学获得博士学位,现任中国科学院空天信息创新研究院助理研究员。主要研究方向为SAR图像智能解译

    祝若鑫(1991–),男,德国工学博士(Dr.-Ing.),西安测绘研究所助理研究员。主要研究方向为社会感知和时空数据挖掘

    孙 显(1981–),男,中国科学院空天信息创新研究院研究员,博士生导师,IEEE高级会员,雷达学报青年编委。主要研究方向为计算机视觉与遥感图像理解

    通讯作者:

    王智睿 zhirui1990@126.com

  • 责任主编:刘畅 Corresponding Editor: LIU Chang
  • 中图分类号: TP753

Supervised Contrastive Learning Regularized High-resolution Synthetic Aperture Radar Building Footprint Generation

Funds: The National Natural Science Foundation of China (62101371, 62076241), Jiangsu Province Science Foundation for Youths (BK20210707)
More Information
  • 摘要: 近年来,高分辨合成孔径雷达(SAR)图像的智能解译技术在城市规划、变化监测等方面得到了广泛应用。不同于光学图像,SAR图像的获取方式、图像中目标的几何结构等因素制约了现有深度学习方法对SAR图像地物目标的解译效果。该文针对高分辨SAR图像城市区域建筑物提取,提出了基于监督对比学习的正则化方法,其主要思想是增强同一类别像素在特征空间中的相似性以及不同类别像素之间的差异性,使得深度学习模型能更加关注SAR图像中建筑物与非建筑物区域在特征空间中的区别,从而提升建筑物识别精度。利用公开的大场景SpaceNet6数据集,通过对比实验,提出的正则化方法,其建筑物提取精度相比于常用的分割方法在不同网络结构下至少提升1%,分割结果证明了该文方法在实际数据上的有效性,可以对复杂场景下的城市建筑物区域进行有效分割。此外,该方法也可以拓展应用于其他SAR图像像素级别的地物分割任务中。

     

  • 图  1  监督对比学习正则化的SAR图像建筑物提取模型示意图

    Figure  1.  Supervised contrastive learning regularized SAR building footprint segmentation model

    图  2  城市地区的多模态遥感图像

    Figure  2.  Multi-modality remote sensing images for urban areas

    图  3  对比学习特征空间中的难易特征

    Figure  3.  The easy and hard query feature vectors for contrastive learning

    图  4  本文所采用的常用的DeepLabV3+与UNet网络结构

    Figure  4.  The CNN architectures of DeepLabV3+ and UNet exploited in this paper

    图  5  SpaceNet6城市地区全极化SAR图像(分辨率:0.5 m)

    Figure  5.  Examples of SpaceNet6 full polarization SAR images (Resolution: 0.5 m)

    图  6  不同方法的建筑物提取结果

    Figure  6.  Different methods for building footprint extraction of SAR images

    图  7  提出算法选取到的难查询像素:第1行为输入SAR图像,第2行为像素预测类别的置信度,颜色越深表示置信度越低

    Figure  7.  Selection of hard query pixels: The first row is the input SAR images and the second row shows the classification confidences

    图  8  DeepLabV3+[ResNet50]模型在有与没有对比损失正则化下得到建筑物区域像素特征之间的相似性直方图

    Figure  8.  The histogram of the feature similarities among the building pixels obtained by the trained DeepLabV3+[ResNet50] models with and without the contrastive loss

    图  9  不同查询及键像素数量对所提出方法的敏感性分析(采用DeepLabV3+网络结构)

    Figure  9.  The sensitivity analysis of the proposed method under different numbers of query and key pixels (The CNN architecture of DeepLabV3+ is adopted)

    图  10  所提出方法得到的大范围城市地区建筑物提取结果(DeepLabV3+[ResNet50])

    Figure  10.  The large-scale urban building extraction result based on the proposed method (DeepLabV3+[ResNet50])

    表  1  训练过程采用的数据增强方法

    Table  1.   Adopted data augmentation methods for training

    数据增强数值
    随机裁剪512×512
    水平翻转0.5
    归一化均值:128 方差:32
    下载: 导出CSV

    表  2  训练过程中的参数设定

    Table  2.   Other parameters for training

    参数数值
    Mq256
    Mk512
    $ \tau $0.1
    D128
    $ \delta $0.97
    Learning rate 1×10–3
    Batch size16
    Epoch200
    下载: 导出CSV

    表  3  不同网络模型及损失函数下的建筑物提取性能比较(单位:%)

    Table  3.   Performance comparison of the building segmentation based on different methods (Unit: %)

    网络模型损失函数IoUDicePrecisionRecall
    U-Net[15]Focal+Dice34.98[1.48]51.81[1.61]62.30[1.50]44.42[2.38]
    DeepLabV3+[ResNet34]Focal+Dice48.40[0.14]65.23[0.13]76.95[0.37]56.61[0.03]
    DeepLabV3+[ResNet34]Focal+Dice+CL49.81[0.31]66.50[0.28]78.32[0.48]57.78[0.68]
    DeepLabV3+[ResNet50]
    DeepLabV3+[ResNet50]
    Focal+Dice
    Focal+Dice+CL
    48.72[0.48]
    50.15[0.62]
    65.51[0.43]
    66.79[0.54]
    76.55[0.31]
    78.30[0.48]
    57.26[0.49]
    58.25[0.87]
    U-Net[ResNet34]Focal+Dice49.09[0.12]65.85[0.11]77.36[0.13]57.33[0.20]
    U-Net[ResNet34]Focal+Dice+CL50.62[0.58]67.21[0.52]76.32[0.25]60.05[0.97]
    U-Net[ResNet50]Focal+Dice49.24[0.30]65.99[0.27]77.37[0.23]57.53[0.34]
    U-Net[ResNet50]Focal+Dice+CL50.99[1.18]67.53[1.03]78.21[1.32]59.50[2.36]
    下载: 导出CSV
  • [1] 徐丰, 王海鹏, 金亚秋. 深度学习在SAR目标识别与地物分类中的应用[J]. 雷达学报, 2017, 6(2): 136–148. doi: 10.12000/JR16130

    XU Feng, WANG Haipeng, and JIN Yaqiu. Deep learning as applied in SAR target recognition and terrain classification[J]. Journal of Radars, 2017, 6(2): 136–148. doi: 10.12000/JR16130
    [2] 王雪松, 陈思伟. 合成孔径雷达极化成像解译识别技术的进展与展望[J]. 雷达学报, 2020, 9(2): 259–276. doi: 10.12000/JR19109

    WANG Xuesong and CHEN Siwei. Polarimetric synthetic aperture radar interpretation and recognition: Advances and perspectives[J]. Journal of Radars, 2020, 9(2): 259–276. doi: 10.12000/JR19109
    [3] 丁赤飚, 仇晓兰, 徐丰, 等. 合成孔径雷达三维成像——从层析、阵列到微波视觉[J]. 雷达学报, 2019, 8(6): 693–709. doi: 10.12000/JR19090

    DING Chibiao, QIU Xiaolan, XU Feng, et al. Synthetic aperture radar three-dimensional imaging—from TomoSAR and array InSAR to microwave vision[J]. Journal of Radars, 2019, 8(6): 693–709. doi: 10.12000/JR19090
    [4] 李宁, 牛世林. 基于局部超分辨重建的高精度SAR图像水域分割方法[J]. 雷达学报, 2020, 9(1): 174–184. doi: 10.12000/JR19096

    LI Ning and NIU Shilin. High-precision water segmentation from synthetic aperture radar images based on local super-resolution restoration technology[J]. Journal of Radars, 2020, 9(1): 174–184. doi: 10.12000/JR19096
    [5] ZHAO Lingjun, ZHOU Xiaoguang, and KUANG Gangyao. Building detection from urban SAR image using building characteristics and contextual information[J]. EURASIP Journal on Advances in Signal Processing, 2013, 2013: 56. doi: 10.1186/1687-6180-2013-56
    [6] TUPIN F and ROUX M. Detection of building outlines based on the fusion of SAR and optical features[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2003, 58(1/2): 71–82.
    [7] XU Feng and JIN Yaqin. Automatic reconstruction of building objects from multiaspect meter-resolution SAR images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2007, 45(7): 2336–2353. doi: 10.1109/TGRS.2007.896614
    [8] MICHAELSEN E, SOERGEL U, and THOENNESSEN U. Perceptual grouping for automatic detection of man-made structures in high-resolution SAR data[J]. Pattern Recognition Letters, 2006, 27(4): 218–225. doi: 10.1016/j.patrec.2005.08.002
    [9] FERRO A, BRUNNER D, and BRUZZONE L. Automatic detection and reconstruction of building radar footprints from single VHR SAR images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2012, 51(2): 935–952.
    [10] ZHANG Fengli, SHAO Yun, ZHANG Xiao, et al. Building L-shape footprint extraction from high resolution SAR image[C]. 2011 Joint Urban Remote Sensing Event, Munich, Germany, 2011: 273–276.
    [11] WANG Yinghua, TUPIN F, HAN Chongzhao, et al. Building detection from high resolution PolSAR data by combining region and edge information[C]. IGARSS 2008-2008 IEEE International Geoscience and Remote Sensing Symposium, Boston, MA, USA, 2009: IV–153.
    [12] GOODFELLOW I, BENGIO Y, and COURVILLE A. Deep Learning[M]. Cambridge: MIT press, 2016: 1–800.
    [13] WANG Xiaying, CAVIGELLI L, EGGIMANN M, et al. Hr-SAR-NET: A deep neural network for urban scene segmentation from high-resolution SAR data[C]. 2020 IEEE Sensors Applications Symposium (SAS), Kuala Lumpur, Malaysia, 2020: 1–6.
    [14] 杜康宁, 邓云凯, 王宇, 等. 基于多层神经网络的中分辨SAR图像时间序列建筑区域提取[J]. 雷达学报, 2016, 5(4): 410–418. doi: 10.12000/JR16060

    DU Kangning, DENG Yunkai, WANG Yu, et al. Medium resolution SAR image time-series built-up area extraction based on multilayer neural network[J]. Journal of Radars, 2016, 5(4): 410–418. doi: 10.12000/JR16060
    [15] SHERMEYER J, HOGAN D, BROWN J, et al. SpaceNet 6: Multi-sensor all weather mapping dataset[C]. The 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, Seattle, USA, 2020: 768–777.
    [16] SHAHZAD M, MAURER M, FRAUNDORFER F, et al. Buildings detection in VHR SAR images using fully convolution neural networks[J]. IEEE Transactions on Geoscience and Remote Sensing, 2019, 57(2): 1100–1116. doi: 10.1109/TGRS.2018.2864716
    [17] JING Hao, SUN Xian, WANG Zhirui, et al. Fine building segmentation in high-resolution SAR images via selective pyramid dilated network[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2021, 14: 6608–6623. doi: 10.1109/JSTARS.2021.3076085
    [18] CHEN Jiankun, QIU Xiaolan, DING Chibiao, et al. CVCMFF Net: Complex-valued convolutional and multifeature fusion network for building semantic segmentation of InSAR images[J]. IEEE Transactions on Geoscience and Remote Sensing, in press. doi: 10.1109/TGRS.2021.3068124
    [19] SUN Yao, HUA Yuansheng, MOU Lichao, et al. CG-Net: Conditional GIS-Aware network for individual building segmentation in VHR SAR images[J]. IEEE Transactions on Geoscience and Remote Sensing, in press. doi: 10.1109/TGRS.2020.3043089
    [20] CHEN L, ZHU Yukun, PAPANDREOU G, et al. Encoder-decoder with atrous separable convolution for semantic image segmentation[C]. The European Conference on Computer Vision, Munich, Germany, 2018: 833–851.
    [21] CHEN Ting, KORNBLITH S, NOROUZI M, et al. A simple framework for contrastive learning of visual representations[C]. The 37th International Conference on Machine Learning, Virtual Event, 2020: 1597–1607.
    [22] HE Kaiming, FAN Haoqi, WU Yuxin, et al. Momentum contrast for unsupervised visual representation learning[C]. The 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, USA, 2020: 9726–9735.
    [23] KHOSLA P, TETERWAK P, WANG Chen, et al. Supervised contrastive learning[C]. Advances in Neural Information Processing Systems, Virtual, 2020: 18661–18673.
    [24] LIU Shikun, ZHI Shuaifeng, JOHNS E, et al. Bootstrapping semantic segmentation with regional contrast[J]. arXiv: 2104.04465, 2021.
    [25] LIN T Y, GOYAL P, GIRSHICK R, et al. Focal loss for dense object detection[C]. The 2017 IEEE International Conference on Computer Vision, Venice, Italy, 2017: 2999–3007.
    [26] RONNEBERGER O, FISCHER P, and BROX T. U-Net: Convolutional networks for biomedical image segmentation[C]. 18th International Conference on Medical Image Computing and Computer-Assisted Intervention, Munich, Germany, 2015: 234–241.
    [27] HE Kaiming, ZHANG Xiangyu, REN Shaoqing, et al. Deep residual learning for image recognition[C]. The 2016 IEEE Conference on Computer Vision and Pattern Recognition, Caesars Palace, Las Vegas, USA, 2016: 770–778.
  • 加载中
图(10) / 表(3)
计量
  • 文章访问数:  2294
  • HTML全文浏览量:  1128
  • PDF下载量:  225
  • 被引次数: 0
出版历程
  • 收稿日期:  2021-09-07
  • 修回日期:  2021-11-12
  • 网络出版日期:  2021-12-03
  • 刊出日期:  2022-02-28

目录

    /

    返回文章
    返回