基于半监督空间-通道选择性卷积核网络的极化SAR图像地物分类

王睿川; 王岩飞

doi:10.12000/JR21080

基于半监督空间-通道选择性卷积核网络的极化SAR图像地物分类

DOI: 10.12000/JR21080

王睿川^{1, 2},
王岩飞^1, ,

1.
中国科学院空天信息创新研究院北京 100190
2.
中国科学院大学电子电气与通信工程学院北京 101408

基金项目: 国家重点研发计划(2017YFB0503001)

详细信息

作者简介:
王睿川(1994–)，男，四川绵阳人，中国科学院空天信息创新研究院博士研究生，研究方向为SAR和极化SAR图像分割和解译、目标检测和识别等

王岩飞(1963–)，男，辽宁沈阳人，中国科学院空天信息创新研究院研究员，博士生导师，主要研究方向为微波成像雷达理论方法及应用、数字信号处理等

通讯作者:
王岩飞 yfwang@mail.ie.ac.cn

责任主编：邹焕新 Corresponding Editor: ZOU Huanxin
中图分类号: TN958
计量
- 文章访问数: 1642
- HTML全文浏览量: 733
- PDF下载量: 167
- 被引次数: 0
出版历程
- 收稿日期: 2021-06-11
- 修回日期: 2021-06-22
- 网络出版日期: 2021-08-28

Terrain Classification of Polarimetric SAR Images Using Semi-supervised Spatial-channel Selective Kernel Network

WANG Ruichuan^{1, 2},
WANG Yanfei^{1
, ,}

1.
Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100190, China
2.
School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences, Beijing 101408, China

Funds: The National Key Research and Development Program (2017YFB0503001)

More Information

Corresponding author: WANG Yanfei, yfwang@mail.ie.ac.cn

摘要

摘要: 针对极化合成孔径雷达(极化SAR)图像地物分类中标注样本数量少的问题，该文提出一种基于空间-通道选择性卷积核全卷积网络(SCSKFCN)和预选-联合优化半监督学习(SPUO)的极化SAR图像地物分类方法。SCSKFCN通过使用空间和通道注意力机制，对不同感受野的特征进行自适应加权融合，有效提升了模型的分类性能。SPUO能够高效地利用标注样本，挖掘无标注样本中蕴含的信息。它采用K-Wishart距离进行样本预选并生成伪标签，然后在联合优化阶段使用真实标注样本和伪标注样本同时对模型进行优化。在模型优化过程中，SPUO对伪标注样本进行两步验证并筛选可靠的伪标注样本参与优化。实验结果表明，该方法能够在只使用少量标注样本的条件下实现高精度、高效率的极化SAR图像地物分类。
- 极化SAR图像地物分类 /
- 全卷积网络 /
- 注意力机制 /
- 半监督学习 /
- 空间-通道选择性卷积核网络
Abstract: In this paper, a Spatial-Channel Selective Kernel Fully Convolutional Network (SCSKFCN) and a Semi-supervised Preselection-United Optimization (SPUO) method are proposed for polarimetric Synthetic Aperture Radar (SAR) image classification. Integrated with spatial-channel attention mechanism, SCSKFCN adaptively fuses features that have different sizes of reception field, and achieves promising classification performance. SPUO can efficiently extract information contained in unlabeled samples according to annotated samples. It utilizes K-Wishart distance to preselect unlabeled samples for pseudo label generation, and then optimizes SCSKFCN with both labeled and pseudo labeled samples. During the training process of SCSKFCN, a two-step verification mechanism is applied on pseudo labeled samples to reserve reliable samples for united optimization. The experimental results show that the proposed SCSKFCN-SPUO can achieve promising performance and efficiency using limited number of annotated pixels.
- Terrain classification of PolSAR images /
- Fully convolutional network /
- Attention mechanism /
- Semi-supervised learning /
- Spatial-Channel Selective Kernel Network (SCSKCN)

HTML全文

图 1 SKNet模块与空间注意力模块

Figure 1. SKNet module and spatial attention module

下载: 全尺寸图片幻灯片

图 2 空间-通道选择性卷积核(SCSK)单元

Figure 2. Spatial-Channel Selective Kernel (SCSK) unit

下载: 全尺寸图片幻灯片

图 3 空间-通道选择性卷积核全卷积网络(SCSKFCN)的架构

Figure 3. Architecture of Spatial-Channel Selective Kernel Fully Convolutional Network (SCSKFCN)

下载: 全尺寸图片幻灯片

图 4 联合优化运行流程图

Figure 4. Flowchart of united optimization

下载: 全尺寸图片幻灯片

图 5 预选-联合优化半监督学习方法流程图

Figure 5. The framework of semi-supervised preselection and united optimization method

下载: 全尺寸图片幻灯片

图 6 Flevoland图像分类结果图

Figure 6. Classification results of Flevoland image with different methods

下载: 全尺寸图片幻灯片

图 7 Oberpfaffenhofen图像分类结果图

Figure 7. Classification results of Oberpfaffenhofen image with different methods

下载: 全尺寸图片幻灯片

图 8 预测概率验证阈值$\delta $对性能的影响

Figure 8. Impact of different values of $\delta $

下载: 全尺寸图片幻灯片

图 9 不同训练集大小的影响

Figure 9. Impact of different sizes of training set

下载: 全尺寸图片幻灯片

图 10 SCSK单元中卷积核尺寸组合的影响

Figure 10. Impact of different combinations of kernel sizes in SCSK unit

下载: 全尺寸图片幻灯片

图 11 不同结构的SCSK单元的组成部分

Figure 11. Building blocks of different architectures of SCSK unit

下载: 全尺寸图片幻灯片

图 12 不同结构的SCSK单元的影响

Figure 12. Impact of different architectures of SCSK unit

下载: 全尺寸图片幻灯片

图 13 SPUO中使用不同距离参数r的影响

Figure 13. Impact of different values of r in SPUO

下载: 全尺寸图片幻灯片

图 14 SPUO中判断准则使用Wishart距离和K-Wishart距离的影响

Figure 14. Impact of using Wishart distance and K-Wishart distance as criterion in SPUO

下载: 全尺寸图片幻灯片

表 1 Flevoland图像分类结果表(%)

Table 1. Classification accuracy comparison on Flevoland image (%)

Method	1	2	3	4	5	6	7	8	9	10
CNN	99.30	93.36	92.22	91.59	93.56	92.79	95.73	97.93	97.54	98.59
PCN	92.11	95.76	98.71	96.49	91.07	96.27	97.52	96.87	96.13	92.23
R5FCN	99.86	98.32	99.32	96.27	93.03	92.12	96.00	97.45	98.15	98.39
SKFCN	99.89	97.39	99.60	98.55	95.42	97.28	96.81	98.80	97.28	98.98
SCSKFCN	99.96	98.60	99.75	98.74	96.47	97.60	97.73	98.99	97.89	98.79
Proposed	99.98	98.87	99.91	99.44	98.93	98.07	99.03	99.78	97.45	99.17

Method	11	12	13	14	15	OA	Kappa	Training (s)		Test (s)
CNN	92.25	96.69	98.84	96.78	65.42	95.19	94.76	247.7		3.1
PCN	96.62	96.32	99.05	98.06	81.26	96.36	96.04	216.8		1.2
R5FCN	97.12	97.68	99.61	99.08	86.12	97.41	97.18	89.3		0.9
SKFCN	99.15	99.08	99.61	99.46	83.77	98.46	98.32	105.6		1.0
SCSKFCN	99.47	99.43	99.64	99.73	82.65	98.80	98.70	112.4		1.1
Proposed	99.56	99.68	99.89	99.50	90.51	99.24	99.18	(19.1)+132.8		1.1
^*注：Water(1); Stembeans(2); Forest(3); Potatoes(4); Grasses(5); Beet(6); Rapeseed(7); Peas(8); Lucerne(9); Bare soil(10); Wheat2(11); Wheat1(12); Wheat3(13); Barley(14); Building(15); Proposed表示SCSKFCN-SPUO方法

下载: 导出CSV

表 2 Oberpfaffenhofen图像分类结果表(%)

Table 2. Classification accuracy comparison on Oberpfaffenhofen image (%)

Class	CNN	ARCN	R5FCN	SKFCN	SCSKFCN	Proposed
Wood land	95.02	96.13	94.97	96.25	96.68	97.18
Open areas	97.17	97.32	97.51	97.76	97.99	98.80
Built-up areas	86.42	93.92	90.47	94.27	94.64	94.86
OA	94.09	96.25	95.27	96.60	96.90	97.51
Kappa	89.88	93.64	91.95	94.23	94.74	95.76
Training (s)	225.9	189.6	120.1	137.2	159.2	(50.3)+173.6
Test (s)	4.6	1.6	1.5	1.7	1.9	1.9
^*注：Proposed表示SCSKFCN-SPUO方法

下载: 导出CSV

参考文献(35)

[1]	LEE J S and POTTIER E. Polarimetric Radar Imaging: From Basics to Applications[M]. Boca Raton, USA, CRC Press, 2017: 1–10.
[2]	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
[3]	WANG Haipeng, XU Feng and JIN Yaqiu. A review of polsar image classification: From polarimetry to deep learning[C]. IEEE International Geoscience and Remote Sensing Symposium, Yokohama, Japan, 2019: 3189–3192. doi: 10.1109/IGARSS.2019.8899902.
[4]	CLOUDE S R and POTTIER E. An entropy based classification scheme for land applications of polarimetric SAR[J]. IEEE Transactions on Geoscience and Remote Sensing, 1997, 35(1): 68–78. doi: 10.1109/36.551935
[5]	FREEMAN A and DURDEN S L. A three-component scattering model for polarimetric SAR data[J]. IEEE Transactions on Geoscience and Remote Sensing, 1998, 36(3): 963–973. doi: 10.1109/36.673687
[6]	DEMPSTER A P, LAIRD N M, and 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): 1–38. doi: 10.1111/j.2517-6161.1977.tb01600.x
[7]	FUKUDA S and HIROSAWA H. Support vector machine classification of land cover: Application to polarimetric SAR data[C]. IGARSS 2001. Scanning the Present and Resolving the Future. Proceedings. IEEE 2001 International Geoscience and Remote Sensing Symposium, Sydney, Australia, 2001: 187–189. doi: 10.1109/IGARSS.2001.976097.
[8]	ZHANG Lamei, SUN Liangjie, and MOON W M. Polarimetric SAR image classification based on contextual sparse representation[C]. IEEE International Geoscience and Remote Sensing Symposium, Milan, Italy, 2015: 1837–1840. doi: 10.1109/IGARSS.2015.7326149.
[9]	ERSAHIN K, CUMMING I G, and WARD R K. Segmentation and classification of polarimetric SAR data using spectral graph partitioning[J]. IEEE Transactions on Geoscience and Remote Sensing, 2010, 48(1): 164–174. doi: 10.1109/TGRS.2009.2024303
[10]	DU Peijun, SAMAT A, WASKE B, et al. Random forest and rotation forest for fully polarized SAR image classification using polarimetric and spatial features[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2015, 105: 38–53. doi: 10.1016/j.isprsjprs.2015.03.002
[11]	LEE J S, GRUNES M R, POTTIER E, et al. Unsupervised terrain classification preserving polarimetric scattering characteristics[J]. IEEE Transactions on Geoscience and Remote Sensing, 2004, 42(4): 722–731. doi: 10.1109/TGRS.2003.819883
[12]	LEE J S, SCHULER D L, LANG R H, et al. K-distribution for multi-look processed polarimetric SAR imagery[C]. IEEE International Geoscience and Remote Sensing Symposium, Pasadena, USA, 1994: 2179–2181. doi: 10.1109/IGARSS.1994.399685.
[13]	DOULGERIS A P, ANFINSEN S N, and ELTOFT T. Classification with a non-Gaussian model for PolSAR data[J]. IEEE Transactions on Geoscience and Remote Sensing, 2008, 46(10): 2999–3009. doi: 10.1109/TGRS.2008.923025
[14]	DOULGERIS A P. An automatic U-distribution and Markov random field segmentation algorithm for PolSAR images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2015, 53(4): 1819–1827. doi: 10.1109/TGRS.2014.2349575
[15]	LECUN Y, BOTTOU L, BENGIO Y, et al. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 1998, 86(11): 2278–2324. doi: 10.1109/5.726791
[16]	SHELHAMER E, LONG J, and DARRELL T. Fully convolutional networks for semantic segmentation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(4): 640–651. doi: 10.1109/TPAMI.2016.2572683
[17]	ZHOU Yu, WANG Haipeng, XU Feng, et al. Polarimetric SAR image classification using deep convolutional neural networks[J]. IEEE Geoscience and Remote Sensing Letters, 2016, 13(12): 1935–1939. doi: 10.1109/LGRS.2016.2618840
[18]	CHEN Siwei and TAO Chensong. PolSAR image classification using polarimetric-feature-driven deep convolutional neural network[J]. IEEE Geoscience and Remote Sensing Letters, 2018, 15(4): 627–631. doi: 10.1109/LGRS.2018.2799877
[19]	CHEN Siwei, LI Yongzhen, DAI Dahai, et al. Uniform polarimetric matrix rotation theory[C]. IEEE International Geoscience and Remote Sensing Symposium, Melbourne, Australia, 2013: 4166–4169. doi: 10.1109/IGARSS.2013.6723751.
[20]	MOHAMMADIMANESH F, SALEHI B, MAHDIANPARI M, et al. A new fully convolutional neural network for semantic segmentation of polarimetric SAR imagery in complex land cover ecosystem[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2019, 151: 223–236. doi: 10.1016/j.isprsjprs.2019.03.015
[21]	LIU Xu, JIAO Licheng, TANG Xu, et al. Polarimetric convolutional network for PolSAR image classification[J]. IEEE Transactions on Geoscience and Remote Sensing, 2019, 57(5): 3040–3054. doi: 10.1109/TGRS.2018.2879984
[22]	LI Yangyang, CHEN Yanqiao, LIU Guangyuan, et al. A novel deep fully convolutional network for PolSAR image classification[J]. Remote Sensing, 2018, 10(12): 1984. doi: 10.3390/rs10121984
[23]	CHEN Yanqiao, LI Yangyang, JIAO Licheng, et al. Adversarial reconstruction-classification networks for PolSAR image classification[J]. Remote Sensing, 2019, 11(4): 415. doi: 10.3390/rs11040415
[24]	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
[25]	BI Haixia, SUN Jian, and XU Zongben. A graph-based semisupervised deep learning model for PolSAR image classification[J]. IEEE Transactions on Geoscience and Remote Sensing, 2019, 57(4): 2116–2132. doi: 10.1109/TGRS.2018.2871504
[26]	XIE Wen, MA Gaini, ZHAO Feng, et al. PolSAR image classification via a novel semi-supervised recurrent complex-valued convolution neural network[J]. Neurocomputing, 2020, 388: 255–268. doi: 10.1016/j.neucom.2020.01.020
[27]	HUA Wenqiang, WANG Shuang, XIE Wen, et al. Dual-channel convolutional neural network for polarimetric SAR images classification[C]. IEEE International Geoscience and Remote Sensing Symposium, Yokohama, Japan, 2019: 3201–3204. doi: 10.1109/IGARSS.2019.8899103.
[28]	LI Xiang, WANG Wenhai, HU Xiaolin, et al. Selective kernel networks[C]. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Long Beach, USA, 2019: 510–519. doi: 10.1109/CVPR.2019.00060.
[29]	WOO S, PARK J, LEE J Y, et al. CBAM: Convolutional block attention module[C]. 15th European Conference on Computer Vision-ECCV 2018, Munich, Germany, 2018: 3–19. doi: 10.1007/978-3-030-01234-2_1.
[30]	PARK J, WOO S, LEE J Y, et al. BAM: Bottleneck attention module[J]. arXiv preprint arXiv: 1807.06514, 2018.
[31]	XU Bing, WANG Naiyan, CHEN Tianqi, et al. Empirical evaluation of rectified activations in convolutional network[J]. arXiv preprint arXiv: 1505.00853, 2015.
[32]	MAAS A L, HANNUN A Y, and NG A Y. Rectifier nonlinearities improve neural network acoustic models[C]. The 30th International Conference on Machine Learning, Atlanta, USA, 2013.
[33]	YU F and KOLTUN V. Multi-scale context aggregation by dilated convolutions[J]. arXiv preprint arXiv: 1511.07122, 2015.
[34]	HU Jie, SHEN Li, ALBANIE S, et al. Squeeze-and-excitation networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 42(8): 2011–2023. doi: 10.1109/TPAMI.2019.2913372
[35]	LI Mu. Efficient mini-batch training for stochastic optimization[C]. The 20th ACM Sigkdd International Conference on Knowledge Discovery and Data Mining, Virtual Event, Singapore, 2014: 661–670.