Hyperspectral remote sensing image classification based on semi-supervised conditional random field

Junfeng WU; Zhiguo JIANG; Haopeng ZHANG; Bowen CAI; Penghao LUO

doi:10.11834/jrs.20176121

Views : 0 下载量: 711 CSCD: 6

PDF
Export
Share
Collection
Album

Hyperspectral remote sensing image classification based on semi-supervised conditional random field
Vol. 21, Issue 4, Pages: 588-603(2017)
Received：19 April 2016，

Accepted：10 January 2017，

Published：2017-05
DOI： 10.11834/jrs.20176121
稿件说明：

移动端阅览

吴俊峰, 姜志国, 张浩鹏, 蔡博文, 罗鹏浩. 2017. 半监督条件随机场的高光谱遥感图像分类. 遥感学报, 21(4): 588–603 DOI： 10.11834/jrs.20176121.

Wu J F, Jiang Z G, Zhang H P, Cai B W and Luo P H. 2017. Hyperspectral remote sensing image classification based on semi-supervised conditional random field. Journal of Remote Sensing, 21(4): 588–603 DOI： 10.11834/jrs.20176121.

摘要

善于捕捉空间信息的条件随机场模型虽然已被应用于高光谱遥感图像分类，但条件随机场的性能受到了标注训练样本数量的制约。为解决上述问题，本文提出了一种半监督条件随机场模型用于高光谱遥感图像分类。在该模型中，首先，利用空间-光谱拉普拉斯支持向量机定义关联势函数，以利用未标注样本中包含的信息获取样本类别概率；然后，在交互势函数中嵌入未标注的空间邻域样本，以充分利用空间信息实现对样本类别概率的修正；最后，采用分布式学习策略和平均场完成半监督条件随机场的训练和推断。本文在两个公开的高光谱数据集(Indian Pines数据集，Pavia University数据集)上进行了实验。实验结果表明Kappa系数提升3.94%。

Abstract

Hyperspectral remote sensing image classification is one of the enormous challenges in the field of applied remote sensing. Traditionally

supervised methods

such as Support Vector Machine (SVM)

dominate this area. Especially

Conditional Random Field (CRF) excels in solving this kind of problem in most cases

due to its prominent ability in formulating the spatial relationship. However

CRF suffers from the availability of large amount of labeled samples

which is labor- and time-consuming to obtain in practice. The accuracy tends to decrease dramatically once labeled samples are not adequate or informative enough. To solve the above problem

a semi-supervised CRF model is proposed in this paper. In the semi-supervised CRF model

the association potential is defined as the spatio-spectral Laplacian Support Vector Machine (ssLapSVM)

to exploit the information contained in the unlabeled samples. And the multi-class probability for each sample is obtained by the ssLapSVM with the one-versus-one scheme. In addition

the interaction potential is newly designed by introducing a weight into the Potts model. Note that

in the classification of hyperspectral remote sensing with limited labeled samples

unlabeled neighbors of one labeled samples may often exist. Thus

the labels of these unlabeled neighbors are assigned based on maximum probability acquired by the ssLapSVM

and use the maximum probability as a weight. In the training phrase

the optimal parameters in the association potential

i.e. ssLapSVM

is firstly trained

and then the whole semi-supervised CRF model is trained to get the optimal parameters in the interaction potential. In the inference phrase

mean-field is adopted to find the optimal label configuration over the testing set. The performance of the proposed semi-supervised CRF model is evaluated on two well-known benchmarks

i.e. Indian Pines scene and Pavia University scene. The objective comparison experiments are carried out among some state-of-the-art methods in terms of kappa statistic. On both Indian Pines (IP) scene and Pavia University (PU) scene

the proposed method can exhibits completely better performance

improve by 4.94%@IP and 3.28%@PU

respectively. In addition

the kappa of the proposed method rises with the increase of the number of labeled training samples. And in most cases

the proposed method shows better performance than other contrast methods under the case of the same training labeled samples. With the increase of trade-off coefficient

kappa statistic rise first and tend to steady

and then degrades dramatically. The proposed method also shows better performance when the larger scope of samples participating in the construction of the interaction potential. In this paper

we have developed a semi-supervised CRF to address the problem of hyperspectral image classification. Our method can effectively improve the kappa statistic under limited labeled-training set by newly designed association and interaction potential. Experiments conducted on two well-known hyperspectral datasets demonstrate the effectiveness of the proposed method. And when compared to related semi-supervised algorithms

the proposed method shows its superiority.

关键词

Keywords

references

Ablin R and Sulochana C H. 2013. A survey of hyperspectral image classification in remote sensing. International Journal of Advanced Research in Computer and Communication Engineering, 2(8): 2986–3000

杜培军, 夏俊士, 薛朝辉, 谭琨, 苏红军, 鲍蕊. 2016. 高光谱遥感影像分类研究进展. 遥感学报, 20(2): 236–256

Du P J, Xia J S, Xue C H, Tan K, Su H J and Bao R. 2016. Review of hyperspectral remote sensing image classification. Journal of Remote Sensing, 20(2): 236–256

Espínola M, Piedra-Fernández J A, Ayala R, Iribarne L and Wang J Z. 2015. Contextual and hierarchical classification of satellite images based on cellular automata. IEEE Transactions on Geoscience and Remote Sensing, 53(2): 795–809

Fauvel M, Benediktsson J A, Chanussot J and Sveinsson J R. 2008. Spectral and spatial classification of hyperspectral data using SVMs and morphological profiles. IEEE Transactions on Geoscience and Remote Sensing, 46(11): 3804–3814

Geng B, Tao D C, Xu C, Yang L J and Hua X S. 2012. Ensemble manifold regularization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(6): 1227–1233

Gomez-Chova L, Camps-Valls G, Munoz-Mari J and Calpe J. 2008. Semisupervised image classification with laplacian support vector machines. IEEE Geoscience and Remote Sensing Letters, 5(3): 336–340

Gu Y F and Feng K. 2013. Optimized laplacian SVM with distance metric learning for hyperspectral image classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 6(3): 1109–1117

Kae A, Sohn K, Lee H and Learned-Miller E. 2013. Augmenting CRFs with boltzmann machine shape priors for image labeling//Proceeding of 2013 IEEE Conference on Computer Vision and Pattern Recognition. Portland, OR, USA: IEEE: 2019–2026 [DOI: 10.1109/CVPR.2013.263]

Lafferty J D, McCallum A and Pereira F C N. 2001. Conditional random fields: probabilistic models for segmenting and labeling sequence data//Proceedings of 18th International Conference on Machine Learning (ICML). San Francisco, CA, USA: Morgan Kaufmann Publishers Inc.: 282–289

李祖传, 马建文, 张睿, 李利伟. 2011. 利用SVM-CRF进行高光谱遥感数据分类. 武汉大学学报(信息科学版), 36(3): 306–310

Li Z C, Ma J W, Zhang R and Li L W. 2011. Classifying hyperspectral data using support vector machine conditional random field. Geomatics and Information Science of Wuhan University, 36(3): 306–310

Melgani F and Bruzzone L. 2004. Classification of hyperspectral remote sensing images with support vector machines. IEEE Transactions on Geoscience and Remote Sensing, 42(8): 1778–1790

Sun L, Wu Z B, Liu J J, Xiao L and Wei Z H. 2015. Supervised spectral-spatial hyperspectral image classification with weighted markov random fields. IEEE Transactions on Geoscience and Remote Sensing, 53(3): 1490–1503

童庆禧, 张兵, 郑兰芬. 2006. 高光谱遥感: 原理、技术与应用. 北京: 高等教育出版社: 1–2

Tong Q X, Zhang B and Zheng L F. 2006. Hyperspectral Remote Sensing: the Principle, Technology and Application. Beijing: Higher Education Press: 1–2

王立国, 杨月霜, 刘丹凤. 2016. 基于改进三重训练算法的高光谱图像半监督分类. 哈尔滨工程大学学报, 37(6): 849–854

Wang L G, Yang Y S and Liu D F. 2016. Semi-supervised classification for hyperspectral image based on improved tri-training method. Journal of Harbin Engineering University, 37(6): 849–854

Xia J S, Chanussot J, Du P J and He X Y. 2014. (Semi-) supervised probabilistic principal component analysis for hyperspectral remote sensing image classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 7(6): 2224–2236

杨国鹏, 余旭初, 陈伟, 刘伟. 2008. 基于核Fisher判别分析的高光谱遥感影像分类. 遥感学报, 12(4): 579–585

Yang G P, Yu X C, Chen W and Liu W. 2008. Hyperspectral remote sensing image classification based on kernel fisher discriminant analysis. Journal of Remote Sensing, 12(4): 579–585

Yang L X, Yang S Y, Jin P L and Zhang R. 2014. Semi-supervised hyperspectral image classification using Spatio-spectral Laplacian support vector machine. IEEE Geoscience and Remote Sensing Letters, 11(3): 651–655

Zhong P and Wang R S. 2008. Learning sparse CRFs for feature selection and classification of hyperspectral imagery. IEEE Transactions on Geoscience and Remote Sensing, 46(12): 4186–4197

Zhong P and Wang R S. 2010. Learning conditional random fields for classification of hyperspectral images. IEEE Transactions on Image Processing, 19(7): 1890–1907

Zhong P and Wang R S. 2014. Jointly learning the hybrid CRF and MLR model for simultaneous denoising and classification of hyperspectral imagery. IEEE Transactions on Neural Networks and Learning Systems, 25(7): 1319–1334

Zhong Y F, Lin X M and Zhang L P. 2014. A support vector conditional random fields classifier with a mahalanobis distance boundary constraint for high spatial resolution remote sensing imagery. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 7(4): 1314–1330

Alert me when the article has been cited

提交

Hyperspectral image classification and application based on relevance vector machine

Integration of a spatial-spectral structure with improved local density for hyperspectral image band selection

Information extraction and system construction of digital forest system based on multisource remote sensing data

In-orbit application testing of 5 m Optical Satellite 02 （ZY1-02E）

Class-independent domain adaptation for hyperspectral image classification