Multivariate Spectral Analysis to Extract Materials from Multispectral Data

摘要

This effort investigates various multivariate analysis techniques forclassification/identification to extract natural and manmade features reliably from broad-band spectral imaging data/multispectral imagery. An enhanced Bayesian method is proposed and is demonstrated to exhibit increased accuracy over three conventional supervised classifiers. Broad-band spectral properties of various materials are examined and the perturbations on spectra of pure materials introduced by mixtures are shown. A mixing model that uses multiple linear regression constrained by two physical properties is tested. Bayesian discriminant, Mahalanobis distance, Euclidean distance, Supervised classification, Sub-pixel demixing, Mixture analysis, Linear models, Linear regression.