Geoscience Reference
In-Depth Information
for improving our understanding of organic matter sources and biogeochemical processes
due to their capacity for producing and testing predictive models.
Some technical aspects of developing chemometric models of spectroscopic data have
been discussed, particularly the importance of preprocessing data properly to obtain mean-
ingful multivariate models. Further, the importance of implementing cross validation pro-
cedures to ensure that especially supervised models are capable of making valid predictions
has been discussed.
Most of the methods in this chapter are based on classical statistical estimators such
as means and variances, and make predictions based on classical linear regression using
least-squares estimates. These approaches have the disadvantage of high sensitivity to
atypical observations (outliers) and to small departures from model assumptions. Robust
algorithms for computing PCA, PLS, PCR (Verboven and Hubert,
2005
), PARAFAC
(Engelen et al.,
2009
), and linear regression (McKean,
2004
), employing estimators that
are less sensitive to outliers and small departures in model assumptions, are often inte-
grated in modern statistical software packages or available online as free toolboxes for
MATLAB.
Several techniques helpful in studying CDOM fluorescence data sets have been demon-
strated using a 13-month time series of fluorescence EEMs, DOC, and nutrient measure-
ments obtained from the Horsen's catchment in Denmark. These examples illustrate that
both two- and three-way chemometric models can be useful for variable reduction, visu-
alizing sources of variability in DOM fluorescence data sets, and for predicting relation-
ships among future samples. Using PLS regression, DOC concentrations within Horsen's
catchment streams were predicted with an accuracy comparable to the error associated with
DOC and fluorescence measurements.
Acknowledgments
C. A. Stedmon acknowledges support by the Carlsberg Foundation and the Danish Research
Council (grant #272-07-0485).
Appendix
Comparisons between PARAFAC spectra in the 5-component (Horsens) model from this
study versus PARAFAC spectra in 10 published studies. Horsens components are num-
bered from 1 to 5 corresponding to the order of components (top left to bottom right) in
Figure 10.7
. The number of the corresponding component in the published model is iden-
tified in the upper third of the table. Spectral congruence (Tucker,
1951
) between similar
components in each pair of models is shown separately for the excitation and emission
spectra. Matched components are both visually and statistically similar and are overlaid on
the Horsens spectra in
Figure 10.7
.
