30 August 2013 The Hughes phenomenon in hyperspectral classification based on the ground spectrum of grasslands in the region around Qinghai Lake
Author Affiliations +
Proceedings Volume 8910, International Symposium on Photoelectronic Detection and Imaging 2013: Imaging Spectrometer Technologies and Applications; 89101G (2013) https://doi.org/10.1117/12.2034457
Event: ISPDI 2013 - Fifth International Symposium on Photoelectronic Detection and Imaging, 2013, Beijing, China
Abstract
Hyperspectral data, consisting of hundreds of spectral bands with a high spectral resolution, enables acquisition of continuous spectral characteristic curves, and therefore have served as a powerful tool for vegetation classification. The difficulty of using hyperspectral data is that they are usually redundant, strongly correlated and subject to Hughes phenomenon where classification accuracy increases gradually in the beginning as the number of spectral bands or dimensions increases, but decreases dramatically when the band number reaches some value. In recent years,some algorithms have been proposed to overcome the Hughes phenomenon in classification, such as selecting several bands from full bands, PCA- and MNF-based feature transformations. Up to date, however, few studies have been conducted to investigate the turning point of Hughes phenomenon (i.e., the point at which the classification accuracy begins to decline). In this paper, we firstly analyze reasons for occurrence of Hughes phenomenon, and then based on the Mahalanobis classifier, classify the ground spectrum of several grasslands which were recorded in September 2012 using FieldSpec3 spectrometer in the regions around Qinghai Lake,a important pasturing area in the north of China. Before classification, we extract features from hyperspectral data by bands selecting and PCA- based feature transformations, and In the process of classification, we analyze how the correlation coefficient between wavebands, the number of waveband channels and the number of principal components affect the classification result. The results show that Hushes phenomenon may occur when the correlation coefficient between wavebands is greater than 94%,the number of wavebands is greater than 6, or the number of principal components is greater than 6. Best classification result can be achieved (overall accuracy of grasslands 90%) if the number of wavebands equals to 3 (the band positions are 370nm, 509nm and 886nm respectively) or the number of principal components ranges from 4 to 6.
© (2013) COPYRIGHT Society of Photo-Optical Instrumentation Engineers (SPIE). Downloading of the abstract is permitted for personal use only.
Weiwei Ma, Weiwei Ma, Cailan Gong, Cailan Gong, Yong Hu, Yong Hu, Peng Meng, Peng Meng, Feifei Xu, Feifei Xu, } "The Hughes phenomenon in hyperspectral classification based on the ground spectrum of grasslands in the region around Qinghai Lake", Proc. SPIE 8910, International Symposium on Photoelectronic Detection and Imaging 2013: Imaging Spectrometer Technologies and Applications, 89101G (30 August 2013); doi: 10.1117/12.2034457; https://doi.org/10.1117/12.2034457
PROCEEDINGS
11 PAGES


SHARE
Back to Top