Abstract
Class Separability Improvement Using Wavelet Transform
Xiuping JIA
Australia
Email: x-jia@adfa.edu.au
Dr Unsw
Abstract
Hyperspectral images recorded from an airborne or spaceborne sensor contain rich information on various ground cover types. However, how to extract all the information out of the data is challenging due to two major difficulties. One is that class signatures can be very similar to each other in the original spectral measurements and class separability need to be enhanced in order to perform a reliable image classification. The other is the high dimensionality needs to be reduced to over come Hughes phenomenon.
Wavelet transform has been investigated to represent class spectral signatures. Multi- level wavelet decomposition with different wavelet functions has been examined to generate distinct wavelet features that can enhance class separability. Feature selection is guided by the Bhattacharyya distance measure which is based on a Gaussian distribution model. The reduced dimensionality makes the class training feasible and reliable. The tests were conducted with HyMap data sets and demonstrated significant improvement in classification performance using the class signatures in wavelet domain.