综合聚类和上下文特征的高光谱影像分类

鲍蕊; 薛朝辉; 张像源; 苏红军; 杜培军

doi:10.13203/j.whugis20150043

综合聚类和上下文特征的高光谱影像分类

鲍蕊^{1, 2, 3,},
薛朝辉^{2, 3},
张像源¹,
苏红军⁴,
杜培军^{2, 3, ,}

1.
天津市地质调查研究院, 天津, 300191
2.
南京大学卫星测绘技术与应用国家测绘地理信息局重点实验室, 江苏南京, 21002
3.
南京大学江苏省地理信息技术重点实验室, 江苏南京, 210023
4.
河海大学地球科学与工程学院, 江苏南京, 210098

基金项目:

江苏省杰出青-基金 BK2012018

国家重大科学仪器设备开发专项 012YQ050250

详细信息

作者简介:
鲍蕊, 硕士, 主要从事高光谱遥感影像分类的理论与方法研究。baoruijiayou@163.com

通讯作者:
杜培军, 博士, 教授。dupjrs@126.com

中图分类号: TP751
计量
- 文章访问数: 1805
- HTML全文浏览量: 166
- PDF下载量: 531
出版历程
- 收稿日期: 2015-09-17
- 发布日期: 2017-07-04

Classification Merged with Clustering and Context for Hyperspectral Imagery

BAO Rui^{1, 2, 3,},
XUE Zhaohui^{2, 3},
ZHANG Xiangyuan¹,
SU Hongjun⁴,
DU Peijun^{2, 3, ,}

1.
Tianjin Institute of Geological Survey, Tianjin 300191, China
2.
Key Laboratory for Satellite Mapping Technology and Applications of National Administration of Surveying, Mapping and Geoinformation of China, Nanjing University, Nanjing 210023, China
3.
Jiangsu Provincial Key Laboratory of Geographic Information Science and Technology, Nanjing University, Nanjing 210023, China
4.
School of Earth Sciences and Engineering, Hohai University, Nanjing 210098, China

Funds:

The Jiangsu province Science Fund for Distinguished Young Scholars BK2012018

the National Key Scientific Instrument and Equipment Development Project 012YQ050250

More Information

Author Bio:
BAO Rui, master, specializes in hyperspectral imagery. E-mail:baoruijiayou@163.com

Corresponding author:
DU Peijun, PhD, professor. E-mail:dupjrs@126.com

摘要

摘要: 常规高光谱影像逐像素分类往往没有考虑空间相关性，分类结果未体现地物的空间关联和分布特征。为了在分类中充分利用空间特征，利用聚类信息并结合隐马尔可夫随机场模型讨论了高光谱遥感影像光谱-空间分类方法。首先，在不同特征提取方法（最小噪声分离、独立成分分析和主成分分析）下，使用不同聚类方法（k-均值、迭代自组织分析算法和模糊c-均值算法）借助隐马尔可夫随机场获取优化的分割图；然后，采用4连通区域标记法对分割区域标记生成图像对象，并根据支持向量机的逐像素分类结果采用多数投票法对图像对象进行分类；最后，借助凹槽窗口邻域滤波技术改进分类结果，削弱“椒盐”现象。该方法综合了监督分类和非监督分类的优势，通过聚类引入地物空间相关性信息，通过隐马尔可夫随机场引入上下文特征，较好地弥补了单纯基于光谱信息分类的不足。
- 聚类 /
- 隐马尔可夫随机场 /
- 支持向量机 /
- 高光谱影像 /
- 光谱-空间分类 /
- 多数投票
Abstract: The traditional pixel-wised classification methods for hyperspectral image (HIS) only consider spectral information while ignoring the spatial information, resulting in a big limit of classification performance. Clustering which could assemble pixels similar in spectral features into spatial adjacent clusters, thus effectively express similarity and spatial correlation of adjacent pixels. In order to take full advantages of spatial correlation, this paper explore a spectral-spatial classification method for HSI merged with clustering and context. Firstly, under condition of different feature extraction(MNF, ICA and PCA), different clustering methods(k-means, ISODATA and FCM) are used in hidden markov random field to obtain optimized segmentation map containing context features; secondly, the regions in the segmentation map are labeled by using a four-connected neighborhood labeling method to generate image objects, and a majority voting method is used to classify the objects based on the initial classification map derived from support vector machine (SVM) optimized by particle swarm optimization (PSO). Finally, a Chamfer neighborhood filtering technique is used to regularize the classification map, which partially reduces the noise. This method utilizing spatial information from clustering and introducing context features from HMRF takes advantage of supervised classification and unsupervised classification to gain noise reduction, high-accuracy and high homogeneity, which makes up for the inadequacy of the classification based only on spectral information. Experiment on ROSIS data set and AVIRIS data set respectively illustrate that the method can obtain better performance in terms of classification. The overall accuracy of ROSIS data set reaches to 98.53%, 5.01% higher than that obtained by SVM. Meanwhile the overall accuracy of AVIRIS data set climbs to 91.97%, 7.01% higher than SVM result. We also find that different feature extraction and different clustering will influence the spectral-spatial method using HMRF with edge-protection.
- clustering /
- hidden Markov random field (HMRF) /
- support vector machines (SVMs) /
- hyperspectral imagery /
- spectral-spatial classification /
- majority voting

HTML全文

致谢: 感谢冰岛大学Benediktsson教授和Ghamisi博士对本文提出的宝贵意见。

图 1 逐对象多数投票过程

Figure 1. Graph of Majority Voting for Every Object

下载: 全尺寸图片幻灯片

图 2 凹槽邻域窗口

Figure 2. Chamfer Neighborhoods

下载: 全尺寸图片幻灯片

图 3 ROSIS University of Pavia数据

Figure 3. ROSIS University of Pavia Data Set

下载: 全尺寸图片幻灯片

图 4 AVIRIS Indian Pines数据

Figure 4. AVIRIS Indian Pines Data Set

下载: 全尺寸图片幻灯片

图 5 ROSIS数据边缘检测结果

Figure 5. Edge Detection Result for ROSIS

下载: 全尺寸图片幻灯片

图 6 ROSIS数据分类结果

Figure 6. Classification Maps of ROSIS Data Set

下载: 全尺寸图片幻灯片

图 7 AVIRIS数据分类结果

Figure 7. Classification Maps of AVIRIS Data Set

下载: 全尺寸图片幻灯片

表 1 分类方法的简写

Table 1 Abbreviations of Classification Method

分类方法	简记方式
K-Means +SVM	K-SVM
ISODATA+SVM	I-SVM
FCM+SVM	F-SVM
KM +MNF+HMRF +SVM	KHM-SVM
ISODATA+MNF+HMRF+SVM	IHM-SVM
FCM+MNF+HMRF+SVM	FHM-SVM
KM+ICA+HMRF+SVM	KHI-SVM
ISODATA+ICA+HMRF+SVM	IHI-SVM
FCM+ICA+HMRF+SVM	FHI-SVM
KM+PCA+HMRF+SVM	KHP-SVM
ISODATA+PCA+HMRF+SVM	IHP-SVM
FCM+PCA+HMRF+SVM	FHP-SVM

下载: 导出CSV

表 2 不同分类方法对ROSIS得到的分类结果

Table 2 Classification Results Obtained by Different Methods for the ROSIS Data Set

分类方法	总体精度/%		平均精度/%		Kappa系数
分类方法	No-PR	PR	No-PR	PR	No-PR	PR
SVM	93.52	95.43	91.99	94.65	0.914	0.952
K-SVM	96.06	97.02	94.28	95.33	0.948	0.960
I-SVM	97.61	98.3	97.22	98.00	0.968	0.977
F-SVM	96.32	97.41	95.51	96.92	0.951	0.966
KHM-SVM	95.78	96.66	94.41	95.21	0.944	0.956
IHM-SVM	97.99	98.53	97.77	98.30	0.973	0.981
FHM-SVM	95.32	96.43	93.82	95.05	0.938	0.953
KHI-SVM	95.53	95.80	92.60	92.99	0.940	0.944
IHI-SVM	96.80	97.14	95.52	95.96	0.957	0.962
FHI-SVM	95.42	96.35	93.02	94.28	0.939	0.951
KHP-SVM	96.8	97.14	95.52	95.96	0.957	0.961
IHP-SVM	95.29	96.09	94.41	95.29	0.937	0.948
FHP-SVM	94.51	95.81	93.54	94.84	0.927	0.943

下载: 导出CSV

表 3 不同分类方法对AVIRIS得到的分类结果

Table 3 Classification Results Obtained by Different Methods for the AVIRIS Data Set

序号	地物名称	训练数量	测试数量	K- SVM	I- SVM	F- SVM	KHM- SVM	IHM- SVM	FHM- SVM	KHI- SVM	IHI- SVM	FHI- SVM	KHP- SVM	IHP- SVM	FHP- SVM
1	苜蓿	5	49	83.93	89.81	98.91	86.77	85.42	71.00	100.00	92.39	92.39	61.96	83.70	82.61
2	非耕犁玉米	143	1 291	87.58	81.32	89.11	89.07	89.37	87.94	79.62	82.95	90.54	84.45	87.18	91.17
3	少耕犁玉米	83	751	85.84	83.97	80.52	87.54	90.11	80.16	74.70	85.95	87.48	74.45	77.83	80.41
4	玉米	23	211	96.71	94.77	88.65	88.40	88.42	86.91	89.16	89.84	85.96	83.34	84.81	85.09
5	草地/牧草	49	448	91.33	91.16	94.30	93.39	90.59	96.00	90.36	92.22	93.28	94.10	93.10	94.87
6	草地/树木	74	673	97.58	97.29	97.73	97.12	95.95	97.65	94.46	93.91	96.89	98.17	94.70	97.66
7	草地/修剪牧草	2	24	96.00	96.00	96.67	98.00	94.00	90.86	94.77	77.14	100.00	90.86	92.72	92.72
8	干草堆	48	441	99.19	99.65	100.00	99.53	98.72	97.12	99.37	99.06	99.58	96.68	98.47	97.61
9	燕麦	2	18	91.86	90.91	84.13	82.64	62.50	64.17	60.00	69.06	92.50	78.10	73.33	85.00
10	非耕犁大豆	96	872	86.71	85.16	84.81	92.38	90.86	88.43	85.61	81.29	85.92	87.26	87.90	88.01
11	少耕犁大豆	246	2 222	90.20	89.66	89.86	92.71	93.31	90.58	89.98	91.26	91.07	91.12	90.84	91.17
12	清理过的大豆地	61	553	87.41	84.88	89.87	87.73	88.56	89.75	90.64	87.33	87.92	90.37	88.29	91.69
13	小麦	21	191	97.60	98.38	99.02	98.09	98.91	99.02	96.56	95.06	97.80	99.03	95.33	99.27
14	木材	129	1 165	95.68	95.16	95.54	95.38	95.51	96.11	94.94	94.69	96.01	95.70	95.02	95.25
15	建筑-草地- 树木-机器	38	342	78.86	76.70	79.91	76.90	81.96	80.25	73.00	72.94	73.72	76.05	71.63	75.17
16	石钢塔	9	86	97.60	98.81	97.33	79.76	89.88	82.55	89.37	80.79	78.90	97.94	93.46	97.94
OA/%				90.52	88.93	90.31	91.66	91.97	90.27	87.70	88.40	90.76	89.16	89.28	90.81
AA/%				91.57	90.85	91.65	88.97	89.63	87.41	87.66	86.62	90.62	87.47	88.02	90.35
Kappa系数				0.892	0.874	0.889	0.905	0.908	0.889	0.859	0.868	0.895	0.876	0.878	0.895

下载: 导出CSV

参考文献(16)

[1]	童庆禧, 张兵, 郑兰芬.高光谱遥感:原理、技术与应用[M].北京:高等教育出版社, 2006 Tong Qingxi, Zhang Bin, Zheng Lanfen. Hyperspectral Remote Sensing:the Principle, Technology and Application[M]. Beijing:Higher Education Press, 2006
[2]	Plaza A, Benediktsson J A, Boardman J W, et al. Recent Advances in Techniques for Hyperspectral Image Processing[J]. Remote Sens Environ, 2009, 113:S110-S122 doi: 10.1016/j.rse.2007.07.028
[3]	Bioucas-Dias J M, Plaza A, Camps-Valls G, et al. Hyperspectral Remote Sensing Data Analysis and Future Challenges[J]. IEEE Geoscience & Remote Sensing Magazine, 2013, 1(2):6-36 http://www.academia.edu/14028682/Hyperspectral_Remote_Sensing_Data_Analysis_and_Future_Challenges
[4]	Camps-Valls G, Tuia D, Bruzzone L, et al. Advances in Hyperspectral Image Classification[J]. IEEE Signal Processing Magazine, 2014, 31(1):45-54 doi: 10.1109/MSP.2013.2279179
[5]	Pal M and Mather P M. Support Vector Machines for Classification in Remote Sensing[J]. International Journal of Remote Sensing, 2005, 26(5):1007-1011 doi: 10.1080/01431160512331314083
[6]	Moser G, Serpico S B. Combining Support Vector Machines and Markov Random Fields in an Integrated Framework for Contextual Image Classification[J].IEEE Transactions on Geoscience & Remote Sensing, 2012, 50(5):1-19 https://www.researchgate.net/publication/235890131_Combining_Support_Vector_Machines_and_Markov_Random_Fields_in_an_Integrated_Framework_for_Contextual_Image_Classification
[7]	Fauvel M, Tarabalka Y, Benediktsson J A, et al. Advances in Spectral-spatial Classification of Hyperspectral Images[J]. Proceedings of the IEEE, 2013, 101(3):652-675 doi: 10.1109/JPROC.2012.2197589
[8]	Plaza A, Martinez P, Perez R, et al. A New Approach to Mixed Pixel Classification of Hyperspectral Imagery based on Extended Morphological Profiles[J]. Pattern Recognition, 2004, 37(6):1097-1116 doi: 10.1016/j.patcog.2004.01.006
[9]	Borhani M, Ghassemian H. Hyperspectral Image Classification Based on Spectral-spatial Features Using Probabilistic SVM and Locally Weighted Markov Random Fields[C]. Intelligent Systems (ICIS), Iran, 2014
[10]	Tarabalka Y, Benediktsson J A, Chanussot J. Spectral-spatial Classification of Hyperspectral Imagery Based on Partitional Clustering Techniques[J].IEEE Transactions on Geoscience and Remote Sensing, 2009, 47(8):2973-2987 doi: 10.1109/TGRS.2009.2016214
[11]	刘国英, 茅力非, 王雷光, 等.基于小波域分层Markov模型的纹理分割[J].武汉大学学报·信息科学版, 2009, 34(5):531-534 http://ch.whu.edu.cn/CN/abstract/abstract1248.shtml Liu Guoying, Mao Lifei, Wang Leiguang, et al. Texture Segmentation Based on a Hierarchical Markov Model in Wavelet Domain[J]. Geomatics and Information Science of Wuhan University, 2009, 34(5):531-534 http://ch.whu.edu.cn/CN/abstract/abstract1248.shtml
[12]	Elliott R J, Aggoun L, Moore J B. Hidden Markov Models:Estimation and Control Stochastic Modelling and Applied Probability[M]. NY:Springer, 2008
[13]	Ghamisi P, Benediktsson J A, Ulfarsson M O. Spectral-spatial Classification of Hyperspectral Images Based on Hidden Markov Random Fields[J]. IEEE Transactions on Geoscience and Remote Sensing, 2014, 52(5):2565-2574 doi: 10.1109/TGRS.2013.2263282
[14]	Gonzalez R C, Woods R E. Digital Image Processing[M]. 2nd ed. Englewood Cliffs, NJ:Prentice-Hall, 2002
[15]	Eddy S R. ProfileHidden Markov Models[J]. Boiinformatics Review, 1998, 14(9):755-763 doi: 10.1093/bioinformatics/14.9.755
[16]	Jain A K, Murty M N, Flynn P J.Data Clustering:A Review[J]. ACM Comput Surv, 1999, 31(3):264-323 doi: 10.1145/331499.331504