Geoscience Reference
In-Depth Information
450
450
0.05
A
400
400
0.04
350
350
0.03
300
300
0.02
250
250
300
400
500
300
400
500
450
450
450
0.01
B
0.02
400
400
400
0
0
−0.02
350
350
350
−0.04
−0.01
300
−0.06
300
300
−0.08
−0.02
250
250
250
300
400
500
300
400
500
300
400
500
Emission wavelength (nm)
Figure 10.12. Latent variables in the PLS prediction of DOC from fluorescence in the Horsens catch-
ment.
(A)
River and WTP model with two latent variables.
(B)
Estuary model with three latent
variables.
signals in the latent variables (
Figure 10.12
), leading to negative regression coefficients
being obtained for wavelengths that actually correlate positively to DOC (Kjeldahla and
Bro,
2010
).
The performance of the PLS regression models can be examined closely in plots of
predicted and measured DOC against time, shown in
Figure 10.13
for two river sites
(R12: agricultural and R13: forested), the WTP site (W16) and an estuarine site (E3). At
each site, temporal trends in measured DOC concentrations are also reflected in predicted
concentrations. Close correspondence between measured and predicted concentrations is
apparent particularly at the river sites (
Figure 10.13A, B
). Predictions at the wastewa-
ter site correspond generally with the DOC data, although not as closely as for the river
sites (
Figure 10.13C
). In the estuary, the pool of nonfluorescent DOM that does not cov-
ary with fluorescent DOM might be expected to be larger due to the decoupling of pro-
cesses that regulate CDOM and DOM production in marine waters (Nelson and Siegel,
2002
). Despite this, not only are the overall trends in DOC concentrations at station E3
adequately modeled, but in many cases the differences in predicted DOC concentrations
obtained from replicate EEMs is smaller than the difference between replicate DOC sam-
ples (
Figure 10.13D
). Overall, the median absolute errors associated with the DOC predic-
tions at the four sites are 4.9%, 3.3%, 11.6%, and 5.9% at sites R12, R13, W16, and E3
respectively. Thus the error associated with the PLS predictions of DOC concentration is
often comparable to, and sometimes better than, the combined (DOC and fluorescence)
method measurement errors.
