Share Email Print
cover

Proceedings Paper

Hyperspectral anomaly detection based on machine learning and building selection graph
Author(s): Yehui Tang; Hanlin Qin; Ying Liang; Hanbing Leng; Zezhao Ju
Format Member Price Non-Member Price
PDF $14.40 $18.00

Paper Abstract

In hyperspectral images, anomaly detection without prior information develops rapidly. Most of the existing methods are based on restrictive assumptions of the background distribution. However, the complexity of the environment makes it hard to meet the assumptions, and it is difficult for a pre-set data model to adapt to a variety of environments. To solve the problem, this paper proposes an anomaly detection method on the foundation of machine learning and graph theory. First, the attributes of vertexes in the graph are set by the reconstruct errors. And then, robust background endmember dictionary and abundance matrix are received by structured sparse representation algorithm. Second, the Euler distances between pixels in lower-dimension are regarded as edge weights in the graph, after the analysis of the low dimensional manifold structure among the hyperspectral data, which is in virtue of manifold learning method. Finally, anomaly pixels are picked up by both vertex attributes and edge weights. The proposed method has higher probability of detection and lower probability of false alarm, which is verified by experiments on real images.

Paper Details

Date Published: 24 October 2017
PDF: 6 pages
Proc. SPIE 10462, AOPC 2017: Optical Sensing and Imaging Technology and Applications, 104625B (24 October 2017); doi: 10.1117/12.2285780
Show Author Affiliations
Yehui Tang, Xidian Univ. (China)
Hanlin Qin, Xidian Univ. (China)
Ying Liang, Xidian Univ. (China)
Hanbing Leng, Xi'an Institute of Optics and Precision Mechanics (China)
Zezhao Ju, Xidian Univ. (China)


Published in SPIE Proceedings Vol. 10462:
AOPC 2017: Optical Sensing and Imaging Technology and Applications
Yadong Jiang; Haimei Gong; Weibiao Chen; Jin Li, Editor(s)

© SPIE. Terms of Use
Back to Top