Modelling Plant Ecology - Environmental Modelling: Finding Simplicity in Complexity - page 207

6

5

4

3

2

1

0

6

5

4

3

2

1

0 *

Jan Apr July Oct Jan Apr July Oct

2002 2003

Time

Figure 12.5 Measured (symbols) and model output from the soil plant atmosphere model (lines) stand-scale sap flow in the control
(upper panel) and rainfall manipulation (lower panel) plot at Caxiuana. Sap flow data scaled to stand-scale from a sample of 12-24
trees. The vertical arrows indicate the beginning of the rainfall manipulation. * denotes a period when the panels were temporarily
removed. Error bars in grey are 1 SD intervals propagated from the confidence in the relationship between tree diameter and sap flow
rate (Reproduced with permission from Fisher et al . (2007), Global Change Biology. Permission Pending. Fisher, R.A., Williams, M.,
Lola da Costa, A. et al . (2007) The response of an Eastern Amazonian rain forest to drought stress: results and modelling analyses
from a through-fall exclusion experiment. Global Change Biology , 13, 1-18).

patterns under conditions of extreme drought. Further-
more, comparisons with 'bottom-up' estimates of carbon
fluxes generated from measurements of growth, litter fall,
ecosystem respiration and tissue turnover by Malhi et al .
(2009), suggest that the estimates of carbon assimila-
tion produced by the model are in very close agreement
with their estimates. These results, which verify both
the outcome of the model, in terms of whole-canopy
water use, and the mechanisms used to generate the pre-
dictions, in terms of leaf-water potential and stomatal
conductance, give us more confidence than we would
have from a purely empirical perspective.
For models of this complexity, quantification of the
actual levels of confidence in model predictions is not
yet available, given the numerous different data streams
and a lack of clarity on how their relative model-data

comparisons should be weighted. For simpler ecosys-
tem models, however, more advanced statistical tech-
niques, referred to as either model-data fusion or data
assimilation have been implemented (see also Chapter 5).
These techniques use the measurement error on data and
the model error generated from ensembles of runs with
varying parameters to estimate levels of confidence in the
real state of a system and then use these to propagate the
levels of error forward in time (Williams et al ., 2005; Fox
et al ., 2009). While these techniques are not yet well estab-
lished for dynamic vegetation models, use of these more
rigorous statistical techniques with complex vegetation
models should eventually provide a more appropriate
framework with which to assess the ability of models and
data to represent the real world efficiently.