A new network based feature-selection approach for hyperspcetral analysis

Hyperspectral imagery contains hundreds of spectral bands, which generates a rather large amount of data, so band selection is often adopted for hyperspectral image analysis. This paper presents a novel approach for unsupervised band selection by transforming the hyperspectral data into complex networks and analyzing the corresponding topological characteristics. The networks´ statistical properties are investigated to evaluate different spectral bands. The objective of the method is to find the bands which can form the most representative network formation. This is a completely new criterion for band selection. Meanwhile, the proposed technique has both an explicit physical meaning and simple process. Experimental results demonstrate that the proposed feature selection approach can acquire better results with respect to the traditional methods.