基于邻域最小生成树的半监督极化SAR图像分类方法

滑文强 王爽 郭岩河 谢雯

滑文强, 王爽, 郭岩河, 等. 基于邻域最小生成树的半监督极化SAR图像分类方法[J]. 雷达学报, 2019, 8(4): 458–470. doi: 10.12000/JR18104
引用本文: 滑文强, 王爽, 郭岩河, 等. 基于邻域最小生成树的半监督极化SAR图像分类方法[J]. 雷达学报, 2019, 8(4): 458–470. doi: 10.12000/JR18104
HUA Wenqiang, WANG Shuang, GUO Yanhe, et al. Semi-supervised PolSAR image classification based on the neighborhood minimum spanning tree[J]. Journal of Radars, 2019, 8(4): 458–470. doi: 10.12000/JR18104
Citation: HUA Wenqiang, WANG Shuang, GUO Yanhe, et al. Semi-supervised PolSAR image classification based on the neighborhood minimum spanning tree[J]. Journal of Radars, 2019, 8(4): 458–470. doi: 10.12000/JR18104

基于邻域最小生成树的半监督极化SAR图像分类方法

doi: 10.12000/JR18104
基金项目: 国家自然科学基金面上项目(61771379),陕西省普通高等学校重点学科专项
详细信息
    作者简介:

    滑文强(1987–),男,陕西西安人,现为西安邮电大学计算机学院讲师,研究方向为极化SAR图像处理。E-mail: huawenqiang2013@163.com

    王爽:王 爽(1978–),女,西安电子科技大学教授,博士生导师,智能信息处理研究所副所长,智能感知与图像理解教育部重点实验室成员,国家“111”计划创新引智基地成员,IEEE会员,IET会员,中国电子学会会员,中国计算机学会会员。主要从事SAR/PolSAR处理与分析、稀疏表示、机器学习等方面的研究工作。E-mail: shwang@mail.xidian.edu.cn

    郭岩河(1990–),男,福建泉州人,西安电子科技大学博士研究生,主要研究方向为极化SAR图像处理、深度学习、机器学习等。E-mail: 1153603266@qq.com

    谢雯:谢 雯(1989–),陕西西安人,现为西安邮电大学通信工程学院讲师,主要研究方向为极化SAR图像处理。E-mail: xiewen236@163.com

    通讯作者:

    滑文强  huawenqiang2013@163.com

  • 中图分类号: TN958

Semi-supervised PolSAR Image Classification Based on the Neighborhood Minimum Spanning Tree

Funds: The National Natural Science Foundation of China (61771379), Shaanxi Key Disciplines of Special Funds Projects
More Information
  • 摘要: 该文针对极化SAR图像分类中只有少量标记样本的问题,提出了一种基于邻域最小生成树的半监督极化SAR图像分类方法。该方法针对极化SAR图像以像素为分类对象的特点,结合自训练方法的思想,利用极化SAR图像像素点的空间信息,提出了基于邻域最小生成树辅助学习的样本选择策略,增加自训练过程中被选择无标记样本的可靠性,扩充标记样本数量,训练更好的分类器。最终用训练好的分类器对极化SAR图像进行测试。对3组真实的极化SAR图像进行测试,实验结果表明,该方法在只有少量标记样本的情况下能获得满意的分类结果,且分类正确率明显优于传统的分类算法。

     

  • 图  1  自训练方法

    Figure  1.  Self-training method

    图  2  极化SAR协方差矩阵中9个元素的灰度值

    Figure  2.  The gray value of 9 elements in PolSAR covariance matrix

    图  3  带权无向图G及其最小生成树

    Figure  3.  Weighted undirected graph G and its minimum spanning tree

    图  4  基于邻域的最小生成树生成过程

    Figure  4.  The spanning process of neighborhood minimum spanning tree

    图  5  基于邻域最小生成树的半监督极化SAR分类方法

    Figure  5.  Semi-supervised PolSAR classification based on the neighborhood minimum spanning tree

    图  6  Flevoland地区AIRSAR L波段数据不同方法的分类结果

    Figure  6.  Classification results of the Flevoland data acquired by AIRSAR

    图  7  Flevoland地区Radarsat-2 C波段数据不同方法的分类结果

    Figure  7.  Classification result of the Flevoland data acquired by Radarsat-2

    图  8  旧金山地区Radarsat-2 C波段数据不同方法的分类结果

    Figure  8.  Classification result of the San Francisco data acquired by Radarsat-2

    图  9  迭代次数对实验结果的影响

    Figure  9.  The effects of number of iterations in the proposed method

    表  1  AIRSAR L波段的Felvoland地区不同分类算法的分类精度(%)

    Table  1.   Classification accuracy of the Flevoland area acquired by AIRSAR L band (%)

    区域方法
    WishartSVMSelf-training本文方法
    Stembeans91.4870.0790.8298.75
    Rapeseed61.8338.0267.1459.58
    Bare soil97.5186.8970.9796.75
    Potatoes79.4758.3880.2781.99
    Beet92.3585.6195.0594.60
    Wheat 267.4371.8067.3989.86
    Peas93.1077.7095.2497.56
    Wheat 382.0882.4294.3397.05
    Lucerne84.5340.7781.6795.06
    Barley81.9698.2998.6298.39
    Wheat81.4668.2885.3485.41
    Grasses66.4965.0381.7580.08
    Forest84.2161.0377.6694.77
    Water46.8565.3269.3993.35
    Building81.7778.912.1885.58
    OA79.4070.3077.1989.92
    下载: 导出CSV

    表  2  AIRSAR L波段的Felvoland 地区不同训练样本的分类结果

    Table  2.   Classification results of the Flevoland area acquired by AIRSAR L band with different number of training samples

    方法训练样本数
    4 6 8 10
    OA (%)KappaOA (%)KappaOA (%)KappaOA (%)Kappa
    Wishart74.620.7215 76.190.7459 78.780.7656 80.260.7831
    SVM56.070.542358.120.561164.420.610270.300.6682
    Self-training63.360.602568.420.656973.890.714677.230.7489
    本文方法79.330.788883.060.809386.900.841689.920.8852
    下载: 导出CSV

    表  3  Radarsat-2 C波段的Felvoland地区不同分类算法的分类精度(%)

    Table  3.   Classification accuracy of the Flevoland area acquired by Radarsat-2 C band (%)

    区域方法
    WishartSVMSelf-training本文方法
    Urban69.6154.7563.9371.44
    Water98.7196.8399.1098.82
    Forest91.6565.2573.8383.63
    Cropland55.2778.9779.2382.24
    OA78.8173.9579.0284.03
    下载: 导出CSV

    表  4  Radarsat-2 C波段的Felvoland 地区不同训练样本的分类结果

    Table  4.   Classification results of the Flevoland area acquired by Radarsat-2 C band with different number of training samples

    方法训练样本数
    4 6 8 10
    OA (%)KappaOA (%)KappaOA (%)KappaOA (%)Kappa
    Wishart69.210.5803 73.650.6239 76.810.6854 78.810.7026
    SVM50.790.415364.790.547170.050.596873.950.6394
    Self-training65.690.523370.410.591174.400.660579.450.7144
    本文方法76.710.676879.290.723582.020.764484.030.7882
    下载: 导出CSV

    表  5  Radarsat-2 C波段的旧金山地区不同分类算法的分类结果(%)

    Table  5.   Classification accuracy of the San Francisco area acquired by radarsat-2 C Band (%)

    区域方法
    WishartSVMSelf-training本文方法
    Water98.7090.0498.0499.92
    Vegetation91.0378.5184.4591.50
    Low-Density Urban81.3042.3170.1875.05
    High-Density Urban42.5877.1533.0168.27
    Developed55.2624.0056.1658.81
    OA73.7762.4068.3778.71
    下载: 导出CSV

    表  6  Radarsat-2 C波段的旧金山地区不同训练样本的分类结果

    Table  6.   Classification results of the San Francisco area acquired by Radarsat-2 C band with different number of training samples

    方法训练样本数
    4 6 8 10
    OA (%)KappaOA (%)KappaOA (%)KappaOA (%)Kappa
    Wishart68.090.5181 70.440.5439 72.490.5867 73.770.6011
    SVM50.240.281751.250.290556.310.362862.400.4342
    Self-training52.340.312658.620.366963.270.435768.420.5308
    本文方法70.870.548273.150.598675.230.628478.710.6852
    下载: 导出CSV
  • [1] NUNZIATA F, MIGLIACCIO M, LI Xiaofeng, et al. Coastline extraction using dual-Polarimetric COSMO-SkyMed PingPong mode SAR data[J]. IEEE Geoscience and Remote Sensing Letters, 2014, 11(1): 104–108. doi: 10.1109/LGRS.2013.2247561
    [2] HE Jinglu, WANG Yinghua, LIU Hongwei, et al. A novel automatic PolSAR ship detection method based on superpixel-level local information measurement[J]. IEEE Geoscience and Remote Sensing Letters, 2018, 15(3): 384–388. doi: 10.1109/LGRS.2017.2789204
    [3] BUONO A, NUNZIATA F, MIGLIACCIO M, et al. Classification of the yellow river delta area using fully polarimetric SAR measurements[J]. International Journal of Remote Sensing, 2017, 38(23): 6714–6734. doi: 10.1080/01431161.2017.1363437
    [4] RATHA D, BHATTACHARYA A, and FRERY A C. Unsupervised classification of PolSAR data using a scattering similarity measure derived from a geodesic distance[J]. IEEE Geoscience and Remote Sensing Letters, 2018, 15(1): 151–155. doi: 10.1109/LGRS.2017.2778749
    [5] LI Dong and ZHANG Yunhua. Adaptive model-based classification of PolSAR data[J]. IEEE Transactions on Geoscience and Remote Sensing, 2018, 56(12): 6940–6955. doi: 10.1109/TGRS.2018.2845944
    [6] HUANG Xiayuan, ZHANG Bo, QIAO Hong, et al. Local discriminant canonical correlation analysis for supervised PolSAR image classification[J]. IEEE Geoscience and Remote Sensing Letters, 2017, 14(11): 2102–2106. doi: 10.1109/LGRS.2017.2752800
    [7] REDOLFI J, SÁNCHEZ J, and FLESIA A G. Fisher vectors for PolSAR image classification[J]. IEEE Geoscience and Remote Sensing Letters, 2017, 14(11): 2057–2061. doi: 10.1109/LGRS.2017.2750800
    [8] LIU Hongying, WANG Yikai, YANG Shuyuan, et al. Large polarimetric SAR data semi-supervised classification with spatial-anchor graph[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2016, 9(4): 1439–1458. doi: 10.1109/JSTARS.2016.2518675
    [9] HUA W Q, WANG S, YANG Zhao et al. Semi-supervised PolSAR image classification based on improved Tri-training[C]. 2017 IEEE International Geoscience and Remote Sensing Symposium, Fort Worth, USA, 2017:3937-3940.
    [10] ROSENBERG C, HEBERT M, and SCHNEIDERMAN H. Semi-supervised self-training of object detection models[C]. Proceedings of the 2005 7th IEEE Workshops on Applications of Computer Vision, Breckenridge, USA, 2005: 29–36.
    [11] BLUM A and MITCHELL T. Combining labeled and unlabeled data with co-training[C]. Proceedings of the 11th Conference on Computational Learning Theory, Madison, USA, 1998: 92–100.
    [12] ZHU Zhihua and LI Ming. Tri-training: Exploiting unlabeled data using three classifiers[J]. IEEE Transactions on Knowledge and Data Engineering, 2005, 17(11): 1529–1541. doi: 10.1109/TKDE.2005.186
    [13] LIU Hongying, WANG Yikai, ZHU Dexiang, et al.. Semi-supervised classification based on anchor-spatial graph for large polarimetric SAR data[C]. Proceedings of 2015 IEEE International Geoscience and Remote Sensing Symposium, Milan, Italy, 2015: 1845–1848.
    [14] LIU Hongying, ZHU Dexiang, YANG Shuyuan, et al. Semisupervised feature extraction with neighborhood constraints for polarimetric SAR classification[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2016, 9(7): 3001–3015. doi: 10.1109/JSTARS.2016.2532922
    [15] WU Wenjin, LI Hailei, ZHANG Lu, et al. High-resolution PolSAR scene classification with pretrained deep convnets and manifold polarimetric parameters[J]. IEEE Transactions on Geoscience and Remote Sensing, 2018, 56(10): 6159–6168. doi: 10.1109/TGRS.2018.2833156
    [16] RASMUS A, VALPOLA H, HONKALA M, et al. Semi-supervised learning with ladder networks[J]. arXiv: 1507.02672, 2015.
    [17] CHENG Yanhua, ZHAO Xin, CAI Rui, et al. Semi-supervised multimodal deep learning for RGB-D object recognition[C]. Proceedings of the 25th International Joint Conference on Artificial Intelligence, New York, USA, 2016: 3345–3351.
    [18] HÄNSCH R and HELLWICH O. Semi-supervised learning for classification of polarimetric SAR-data[C]. Proceedings of 2009 IEEE International Geoscience and Remote Sensing Symposium, Cape Town, South Africa, 2009: 987–990.
    [19] Liu H Y, Wang Y K, Zhua D X et al.. Semi-supervised classification based on anchor-spatial graph for large polarimetric SAR data[C]. 2015 IEEE International Geoscience and Remote Sensing Symposium, Milan, Italy, 2015: 1845-1848.
    [20] HUA Wenqiang, WANG Shuang, LIU Hongying, et al. Semisupervised PolSAR image classification based on improved cotraining[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2017, 10(11): 4971–4986. doi: 10.1109/JSTARS.2017.2728067
    [21] GENG Jie, MA Xiaorui, FAN Jianchao, et al. Semisupervised classification of polarimetric SAR image via superpixel restrained deep neural network[J]. IEEE Geoscience and Remote Sensing Letters, 2018, 15(1): 122–126. doi: 10.1109/LGRS.2017.2777450
    [22] LASZLO M and MUKHERJEE S. Minimum spanning tree partitioning algorithm for microaggregation[J]. IEEE Transactions on Knowledge and Data Engineer, 2005, 17(7): 902–911. doi: 10.1109/TKDE.2005.112
    [23] 王晓东. 计算机算法设计与分析[M]. 第4版, 北京: 电子工业出版社, 2012: 103–104.

    WANG Xiaodong. Design and Analysis of Algorithms[M]. 4th Ed, Beijing: China, Electronic Industry Press, 2002: 103–104.
    [24] LEE J S, GRUNES M R, AINSWORTH T L, et al. Unsupervised classification using polarimetric decomposition and the complex Wishart classifier[J]. IEEE Transactions on Geoscience and Remote Sensing, 1999, 37(5): 2249–2258. doi: 10.1109/36.789621
    [25] LEE J S, GRUNES M R, and DE GRANDI G. Polarimetric SAR speckle filtering and its implication for classification[J]. IEEE Transactions on Geoscience and Remote Sensing, 1999, 37(5): 363–373. doi: 10.1109/36.789635
    [26] LONG Y, and LIU X. SVM lithological classification of PolSAR image in yushigou Area, Qilian Mountain[J]. Scientific Journal of Earth Science, 2013, 3(4): 128–132.
    [27] LEE J S, GRUNES M R, and KWOK R. Classification of multi-look polarimetric SAR imagery based on complex Wishart distribution[J]. International Journal of Remote Sensing, 1994, 15(11): 2299–2311. doi: 10.1080/01431169408954244
  • 加载中
图(9) / 表(6)
计量
  • 文章访问数:  1694
  • HTML全文浏览量:  584
  • PDF下载量:  167
  • 被引次数: 0
出版历程
  • 收稿日期:  2018-12-03
  • 修回日期:  2018-12-28
  • 网络出版日期:  2019-02-19
  • 刊出日期:  2019-08-28

目录

    /

    返回文章
    返回