Hyperspectral Imaging and Feature Mining

Xiuping Jia
UNSW

     Full text: Not available
     Last modified: July 31, 2011

Abstract
As the advance of the technology in sensor design and manufacture, hyperspectral imagers have become available since late of 80s, such as AVIRIS, HyMap, for airborne remote sensing and Hyperion on EO1 for spaceborne remote sensing of environment. A typical hyperspectral data cube is composed of about 100 to 200 spectral measurements in the wavelength range of 400 to 2400 nm for each spatial element of an imaged scene. They provide rich spectral information of the ground cover materials, in general, due to its high spectral resolution and full coverage of the optical reflective wavelength region. However, not all the measurements are important and essential for individual applications where users have their focused culture or natural classes of interest. The use of the unnecessary measurements creates difficulties in machine learning processing and reduce the reliability of thematic mapping when the training sample sizes are small, which is called Hughes phenomenon. Therefore for a given application, there are two issues to address. One is to remove the redundant measurements. The other is to select or derive useful features from the original data.

In this paper, redundant reduction is investigated using an unsupervised statistical method; named Geometric based feature selections (G-FS). The useful features are obtained via first generate more features beyond spectral measures, including spatial texture features, such as contrast, homogeneity and energy, and structure enhanced new features, followed by supervised feature selection methods. An information class separability measure in cluster space is presented in this paper. Experiments are conducted and the results demonstrated that the reduced subset of features could minimize redundancies effectively, enhance class separability and avoid the Hughes phenomenon.