This paper addresses the problem of assessing the robustness with respect t
o change in parameters of an integrated training and classification routine
. Extensions to traditional multivariate statistical methods are applied to
perform the classification. Training sets are grown from one or a few seed
points by a method that ensures spatial and spectral closeness of observat
ions. Spatial closeness is obtained by requiring connectivity. Spectral clo
seness is obtained by excluding observations that have high Mahalanobis dis
tances to the training class mean. The marginal effects of changes in the p
arameters that are input to the seed growing algorithm are evaluated. Initi
ally, the seed is expanded to a small area in order to allow for the estima
tion of a dispersion matrix. This expansion is controlled by upper limits f
or the spatial and Euclidean spectral distances from the seed point. Second
, after this initial expansion, the growing of the training set is controll
ed by an upper limit for the Mahalanobis distance to the current estimate o
f the class centre. Also, the estimates of class centres and dispersion mat
rices may be continuously updated, or the initial estimates may be used. Fi
nally, the effect of the operator's choice of seed among a number of potent
ial seed points is evaluated. An evaluation of the sensitivity of the seed
algorithm with respect to parameter settings is carried out by applying it
to the classification of minerals commonly encountered in siliciclastic or
carbonate rocks from twelve chemical elements mapped from thin sections by
energy-dispersive spectroscopy (EDS) in a scanning electron microscope (SEM
), using a standard quadratic classifier. The performance for each paramete
r setting is measured by the overall misclassification rate on an independe
ntly generated validation set. The integrated training and classification m
ethod is presently used as a routine petrographical analysis method at Nors
k Hydro Research Centre. (C) 2000 Elsevier Science B.V. All rights reserved
.