深度学习在SAR目标识别与地物分类中的应用

徐丰; 王海鹏; 金亚秋

doi:10.12000/JR16130

深度学习在SAR目标识别与地物分类中的应用

doi: 10.12000/JR16130

复旦大学电磁波信息科学教育部重点实验室上海 200433

基金项目: 国家自然科学基金(61571132, 61571134, 61331020)，上海航天科技创新基金

详细信息

作者简介:
徐丰(1982–)，男，浙江东阳人，复旦大学博士学位，教授，复旦大学电磁波信息科学教育部重点实验室副主任，研究方向为SAR图像解译、电磁散射建模、人工智能，兼职： IEEE地球科学与遥感快报副主编、 IEEE地球科学与遥感学会上海分会主席。E-mail: fengxu@fudan.edu.cn

王海鹏(1979–)，男，河南遂平人，复旦大学电磁波信息科学教育部重点实验室副教授，研究方向为雷达系统设计与算法开发、遥感图像处理与信息获取、机器学习与目标识别、智能图像处理等。E-mail: hpwang@fudan.edu.cn

金亚秋(1946–)，男，上海人，美国麻省理工学院博士学位，教授，复旦大学电磁波信息科学教育部重点实验室主任，中国科学院院士，研究方向为复杂自然介质的电磁辐射、散射与传输。E-mail: yqjin@fudan.edu.cn

通讯作者:
徐丰 fengxu@fudan.edu.cn

中图分类号: TN959
计量
- 文章访问数: 9202
- HTML全文浏览量: 3148
- PDF下载量: 2980
- 被引次数: 0
出版历程
- 收稿日期: 2016-11-29
- 修回日期: 2017-03-14
- 网络出版日期: 2017-04-28

Deep Learning as Applied in SAR Target Recognition and Terrain Classification

Key Laboratory for Information Science of Electromagnetic Waves, Fudan University, Shanghai 200433, China

Funds: The National Natural Science Foundation of China (61571132, 61571134, 61331020), The Foundation of Shanghai Aerospace Science and Technology

摘要

摘要:
深度卷积网络等深度学习算法变革了计算机视觉领域，在多种应用上的效果都超过了以往传统图像处理算法。该文简要回顾了将深度学习应用在SAR图像目标识别与地物分类中的工作。利用深度卷积网络从SAR图像中自动学习多层的特征表征，再利用学习到的特征进行目标检测与目标分类。将深度卷积网络应用于SAR目标分类数据集MSTAR上，10类目标平均分类精度达到了99%。针对带相位的极化SAR图像，该文提出了复数深度卷积网络，将该算法应用于全极化SAR图像地物分类，Flevoland 15类地物平均分类精度达到了95%。
- 合成孔径雷达 /
- 深度学习 /
- 自动目标识别 /
- 地物分类
Abstract:
Deep learning such as deep neural networks has revolutionized the computer vision area. Deep learning-based algorithms have surpassed conventional algorithms in terms of performance by a significant margin. This paper reviews our works in the application of deep convolutional neural networks to target recognition and terrain classification using the SAR image. A convolutional neural network is employed to automatically extract a hierarchic feature representation from the data, based on which the target recognition and terrain classification can be conducted. Experimental results on the MSTAR benchmark dataset reveal that deep convolutional network could achieve a state-of-the-art classification accuracy of 99% for the 10-class task. For a polarimetric SAR image classification, we propose complex-valued convolutional neural networks for complex SAR images. This algorithm achieved a state-of-the-art accuracy of 95% for the 15-class task on the Flevoland benchmark dataset.
- Synthetic Aperture Radar (SAR) /
- Deep learning /
- Automatic Target Recognition (ATR) /
- Terrain classification

HTML全文

图 1 深度卷积网络示意图^[9]

Figure 1. Deep convolutional neural networks^[9]

下载: 全尺寸图片幻灯片

图 2 层次化可组合的特性表征能力

Figure 2. Hierarchical compositional feature representation capability

下载: 全尺寸图片幻灯片

图 3 两种激活函数及其梯度比较

Figure 3. Two types of activation function and their gradients

下载: 全尺寸图片幻灯片

图 4 通过ReLU的非线性解决XOR问题

Figure 4. XOR problem solved by using the nonlinearity of ReLU

下载: 全尺寸图片幻灯片

图 5 复数卷积神经网络结构^[10]

Figure 5. Structure of complex-valued convolutional neural network^[10]

下载: 全尺寸图片幻灯片

图 6 网络整体结构，包含5个可训练层。卷积层表示为“conv.(卷积核数量)@(卷积核大小)”

Figure 6. Network architecture, including 5 trainable layers. Convolution layer is denoted as “conv.(kernel depth)@(kernel size)”

下载: 全尺寸图片幻灯片

图 7 10类军事目标示例：光学图像 vs. SAR图像

Figure 7. Ten-class military targets: optical image vs. SAR image

下载: 全尺寸图片幻灯片

图 8 实数卷积神经网络结构

Figure 8. Real-valued convolutional neural network architecture

下载: 全尺寸图片幻灯片

图 9 全极化L-波段旧金山地区AirSAR图像

Figure 9. Fully-polarimetric L-band AirSAR image of San Francisco area

下载: 全尺寸图片幻灯片

图 10 CV-CNN的整体结构

Figure 10. Architecture of CV-CNN

下载: 全尺寸图片幻灯片

图 11 Flevoland数据

Figure 11. Flevoland data

下载: 全尺寸图片幻灯片

图 12 分类结果

Figure 12. Classification result

下载: 全尺寸图片幻灯片

图 13 用于训练的南京某地区ALOS2图像

Figure 13. ALOS2 image of Nanjing used as training data

下载: 全尺寸图片幻灯片

图 14 用于测试的南京另一区域的结果

Figure 14. Results of another region in Nanjing used for testing

下载: 全尺寸图片幻灯片

图 15 用于测试的上海某区域的ALOS图像及其分类结果和对比光学影像

Figure 15. Results of Shanghai area used for testing

下载: 全尺寸图片幻灯片

图 16 数据驱动-模型约束下的SAR智能解译

Figure 16. Data driven-model regularized intelligent SAR imagery interpretation

下载: 全尺寸图片幻灯片

表 1 SAR图像解译与计算机视觉的差异

Table 1. Comparison of SAR imagery interpretation and computer vision

特性	SAR图像解译	计算机视觉
波段	微波	可见光
成像原理	相位相干叠加	能量聚焦叠加
投影方向	距离向-方位角	俯仰角-方位角
典型视角	空天对地观测	第一人称视角
数据构成	幅度相位、多通道、多模式	RGB、视频

下载: 导出CSV

表 2 SOC实验条件下的混淆矩阵

Table 2. Confusion matrix under SOC setting

Class	BMP-2	BTR-70	T-72	BTR-60	2S1	BRDM-2	D7	T-62	ZIL-131	ZSU-234	P_cc (%)
BMP-2	194	0	1	0	1	0	0	0	0	0	98.98
BTR-70	0	195	0	0	0	1	0	0	0	0	99.49
T-72	0	0	196	0	0	0	0	0	0	0	100
BTR-60	1	0	0	188	0	0	0	1	1	4	96.41
2S1	0	0	0	0	269	4	0	0	0	1	98.18
BRDM-2	0	0	0	0	0	272	0	0	0	2	99.27
D7	0	0	0	0	0	0	272	1	1	0	99.27
T-62	0	0	0	0	0	0	0	272	1	0	99.64
ZIL-131	0	0	0	0	0	0	0	0	273	1	99.64
ZSU-234	0	0	0	0	0	0	1	0	0	273	99.64
Total											99.13

下载: 导出CSV

表 3 Flevoland数据分类结果

Table 3. Flevoland result

Class	训练样本数	测试样本数	实数网络准确率	复数网络正确率
Stem beans	1245	282	100	100
Peas	1225	308	99.26	98.38
Forest	1184	295	99.36	99.66
Lucerne	1225	293	93.16	99.66
Wheat	1163	291	97.60	95.88
Beet	1213	287	97.18	99.30
Potatoes	1220	286	89.42	97.20
Bare soil	1236	310	99.36	88.06
Grasses	1202	329	71.43	91.79
Rapeseed	1216	290	87.58	91.72
Barley	1207	311	98.49	99.36
Wheat2	1176	304	83.23	92.43
Wheat3	1224	325	77.61	94.77
Water	1138	284	59.27	96.13
Buildings	126	24	100	100
Total	17000	4219	89.49	95.97

下载: 导出CSV

参考文献(24)

[1]	张红, 王超, 张波, 等. 高分辨率SAR图像目标识别[M]. 北京: 科学出版社, 2009: 4–7. Zhang Hong, Wang Chao, Zhang Bo, et al. High-Resolution SAR Image Target Recognition[M]. Beijing: Science Press, 2009: 4–7.
[2]	Moreira A, Prats-Iraola P, Younis M, et al. A tutorial on synthetic aperture radar[J].IEEE GeoscienceandRemote Sensing Magazine, 2013, 1(1): 6–43.
[3]	程肖. 基于散射中心模型的SAR图像自动目标识别[D]. [硕士论文], 国防科学技术大学, 2009. Cheng Xiao. SAR ATR algorithm based on the scattering center model[D]. [Master dissertation], National University of Defense Technology, 2009.
[4]	Krizhevsky A, Sutskever I, and Hinton G E. ImageNet classification with deep convolutional neural networks[J].Advances in Neural Information Processing Systems, 2012, 25(2): 1097–1105. http://web.cs.ucdavis.edu/~yjlee/teaching/ecs289g-fall2016/yuanzhe.pdf
[5]	He K, Zhang X, Ren S, et al. Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification[C]. IEEE International Conference on Computer Vision, Santiago, Chile, 2015: 1026–1034.
[6]	Szegedy C, Liu W, Jia Y, et al. Going deeper with convolutions[C]. IEEE Conference on Computer Vision and Pattern Recognition, Boston, M. A., USA, 2015: 1–9.
[7]	Large Scale Visual Recognition Challenge 2016 (ILSVRC2016). http://image-net.org/challenges/ LSVRC/2016/results.
[8]	Glorot X, Bordes A, and Bengio Y. Deep sparse rectifier networks[C]. International Conference on Artificial Intelligence and Statistics, Fort Lauderdale, United States, 2011: 315–323.
[9]	Chen S, Wang H, Xu F, et al. Target classification using the deep convolutional networks for SAR images[J].IEEE Transactions on GeoscienceandRemote Sensing, 2016, 54(8): 4806–4817. http://adsabs.harvard.edu/abs/2016ITGRS.54.4806C
[10]	Zhang Z, Wang H, Xu F, et al. Complex-valued convolutional neural network and its application in polarimetric SAR image classification[J].IEEE Transactions on Geoscience &Remote Sensing, in press.
[11]	Hänsch R and Hellwich O. Complex-valued convolutional neural networks for object detection in PolSAR data[C]. European Conference on Synthetic Aperture Radar, Aachen, Germany, 2010: 1–4.
[12]	Lecun Y, Bengio Y, and Hinton G. Deep learning[J].Nature, 2015, 521(7553): 436–444. doi: 10.1038/nature14539
[13]	Zeiler M D and Fergus R. Visualization and understanding convolutional networks[C]. European Conference on Computer Vision, Zurich, Switzerland, 2014: 818–833.
[14]	Srivastava N, Hinton G, Krizhevsky A, et al. Dropout: A simple way to prevent neural networks from overfitting[J].Journal of Machine Learning Research, 2014, 15(1): 1929–1958. https://www.researchgate.net/publication/286794765_Dropout_A_Simple_Way_to_Prevent_Neural_Networks_from_Overfitting
[15]	Keydel E R. MSTAR extended operating conditions: A tutorial[C]. Algorithms for Synthetic Aperture Radar Imagery Ⅲ, Orlando, F. L., USA, 1996: 228–242.
[16]	Kong J A, Swartz A A, Yueh H A, et al. Identification of terrain cover using the optimumterrain classifier[J].Journal of Electromagnetic Waves &Applications, 2012, 2(2): 171–194.
[17]	Lee J S and Grunes M R. Classification of multi-look polarimetric SAR data based on the complex Wishart distribution[J].International Journal of Remote Sensing, 1994, 15(11): 2299–2311. doi: 10.1080/01431169408954244
[18]	Freeman A, Villasenor J, Klein J D, et al. On the use of multi-frequency and polarimetric radar backscatter features for classification of agricultural crops[J].International Journal of Remote Sensing, 1994, 15(9): 1799–1812. doi: 10.1080/01431169408954210
[19]	Kouskoulas Y, Ulaby F T, and Pierce L E. The Bayesian Hierarchical Classifier (BHC) and its application to short vegetation using multi-frequency polarimetric SAR[J].IEEE Transactions on GeoscienceandRemote Sensing, 2004, 42(2): 469–477.
[20]	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 GeoscienceandRemote Sensing, 1999, 37(5): 2249–2258.
[21]	Cloude S R and Pottier E. An entropy based classification scheme for land applications of polarimetric SAR[J].IEEE Transactions on GeoscienceandRemote Sensing, 1997, 35(1): 68–78. http://www.academia.edu/10870525/An_entropy_based_classification_scheme_for_land_applications_of_polarimetric_SAR
[22]	Zhou Y, Wang H, Xu F, 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
[23]	Spaceborne and Airborne POLSAR Images[OL]. Available: http://envisat.esa.int/POLSARpro/datasets.html29.
[24]	Yu P, Qin A K, and Clausi D A. Unsupervised polarimetric SAR image segmentation using region growing with edge penalty[J].IEEE Transactions on GeoscienceandRemote Sensing, 2012, 50(4): 1302–1317.