420 lohn Woods
18.17.2 Features of the ecosystem
The choice offeature is limited by the design ofthe model. For example, a food
chain model, with only one functional group of plankton at each trophic level, contains no infonnation about the seasonal succes sion of species. And a one-dimensional model contains no infonnation about patchiness due to mesoscale
turbulence. Such limitations can in principle be overcome by processing the observations to create products commensurate with the simulation. However such processing is in itself a source of uncertainty. In the first example, it is difficult to recreate the time series of biomass for a single function group of plankton (with the
specification used in the model, where there is no competition with other functional
groups), from observed time series of the abundance of many species, including
some from other functional groups. How, for example, can one extract infonnation
from observations about a functional group like diatoms, the growth of which is
limited in the model by nitrogen, but in the ocean by nitrogen and silicon? In the
second example, it is difficult to establish the time series of phytoplankton biomass
in the water column of a one-dimensional simulation, from samples made in the
ocean where the phytoplankton is patchy. It is not sufficient merely to average the
latter, because the patchiness results from complex interaction between the mesoscale turbulence and the plankton, which leads to significant deviation from onedimensional behaviour. The processes that cause patchiness bias biological production (Martin et al 2001). The demographic attractor of a one-dimensional simulation will not be the same as that computed by horizontal averaging of the plankton
populations in a three-dimensional simulation of a patchy ecosystem.
18.17.3 SeaWIFS ocean colour data
In practice, the uncertainties inherent in any feature described by simulation and
observation are so great that it is rare for the Ecological Turing Test to reject a simulation as being so unrealistic that it could not have occurred in Nature. Only when
the modelling process is severely flawed will there be clear blue water between the
uncertain locations of the simulation and observation in feature space. A good
example is provided by Liu et al (2001) who compared a time series of mixed layer
chlorophyll concentration during the 1999 spring bloom just North of the Azores
(Fig. 18.10). In this case, the WB model driven by Bunker climatology produces
significant differences in the timing and amplitude of the spring bloom from those
extracted from Sea WIFS observations (Bahenfeld et al 2001), after statistical processing to eliminate patchiness. Stimulated by the results of this ETT, the authors
subsequently investigated a number ofpossible deficiencies in the model and in the
data used to integrate it. The first hypothesis was that the optical sub-program in
the WB model caused the trouble, but replacing it with a more ac curate (and more
costly) RTE (Radiative Transfer Equation) code only made the matter worse. However, using a more accurate parametrization of cloud-radiation interaction halved
the gap between observation and simulation. And further research showed that the
remaining gap could be closed by a small number of changes to the model, including optimising the infant mortality parameter (Al Battran et al 1998), improving the
18.17.2 Features of the ecosystem
The choice offeature is limited by the design ofthe model. For example, a food
chain model, with only one functional group of plankton at each trophic level, contains no infonnation about the seasonal succes sion of species. And a one-dimensional model contains no infonnation about patchiness due to mesoscale
turbulence. Such limitations can in principle be overcome by processing the observations to create products commensurate with the simulation. However such processing is in itself a source of uncertainty. In the first example, it is difficult to recreate the time series of biomass for a single function group of plankton (with the
specification used in the model, where there is no competition with other functional
groups), from observed time series of the abundance of many species, including
some from other functional groups. How, for example, can one extract infonnation
from observations about a functional group like diatoms, the growth of which is
limited in the model by nitrogen, but in the ocean by nitrogen and silicon? In the
second example, it is difficult to establish the time series of phytoplankton biomass
in the water column of a one-dimensional simulation, from samples made in the
ocean where the phytoplankton is patchy. It is not sufficient merely to average the
latter, because the patchiness results from complex interaction between the mesoscale turbulence and the plankton, which leads to significant deviation from onedimensional behaviour. The processes that cause patchiness bias biological production (Martin et al 2001). The demographic attractor of a one-dimensional simulation will not be the same as that computed by horizontal averaging of the plankton
populations in a three-dimensional simulation of a patchy ecosystem.
18.17.3 SeaWIFS ocean colour data
In practice, the uncertainties inherent in any feature described by simulation and
observation are so great that it is rare for the Ecological Turing Test to reject a simulation as being so unrealistic that it could not have occurred in Nature. Only when
the modelling process is severely flawed will there be clear blue water between the
uncertain locations of the simulation and observation in feature space. A good
example is provided by Liu et al (2001) who compared a time series of mixed layer
chlorophyll concentration during the 1999 spring bloom just North of the Azores
(Fig. 18.10). In this case, the WB model driven by Bunker climatology produces
significant differences in the timing and amplitude of the spring bloom from those
extracted from Sea WIFS observations (Bahenfeld et al 2001), after statistical processing to eliminate patchiness. Stimulated by the results of this ETT, the authors
subsequently investigated a number ofpossible deficiencies in the model and in the
data used to integrate it. The first hypothesis was that the optical sub-program in
the WB model caused the trouble, but replacing it with a more ac curate (and more
costly) RTE (Radiative Transfer Equation) code only made the matter worse. However, using a more accurate parametrization of cloud-radiation interaction halved
the gap between observation and simulation. And further research showed that the
remaining gap could be closed by a small number of changes to the model, including optimising the infant mortality parameter (Al Battran et al 1998), improving the