Integrated wavelet principal component mapping for unsupervised clustering on near infra-red spectra

Abstract

We introduce a new method of unsupervised cluster exploration and visualization for spectral datasets by integrating the wavelet transform, principal components and Gaussian mixture models. The Bayesian Information Criterion (BIC) and classification uncertainty performance criteria are used to guide an automated search of commonly available wavelets and adaptive wavelets. We demonstrate the effectiveness of the proposed method in elucidating and visualizing unsupervised clusters from near infrared (NIR) spectral datasets. The results show that informative feature extraction can be achieved through both commonly available wavelet bases and adaptive wavelets. However, the features from the adaptive wavelets are more favorable in conjunction with unsupervised Gaussian mixture models through a user specified internal linkage function.