DATA ASSIMILATION IN BIOGEOCHEMICAL MODELLING
543
more degrees of freedom to get closer to the observations. There are
thus two contradictory statements about the required model complexity. This may to some extent result from errors in the physics used to
drive the ecosystem models, but the problem is persistent even when
physical observations are used to provide a physical environment as
alistic as possible in a one-dimensional framework. Another possible
explanation for the apparently contradictory statements about ecosystem model complexity is that the NPZD-type models employed so far
may not have the right structures and hence are inherently inconsistent
with the yet unknown ecological rules of real marine ecosystems. It is
not clear whether more complex function-group type models would, in
this data assimilative respect, perform any better. Although their orderof-magnitude larger number in adjustable parameters should allow for
a much better fit to the data, such a fit would be of little value (and
correspond to overfitting few data points by a high-order polynomial)
for applying such models to other climate conditions unless all of the
model parameters can be constrained by observations.
A possible strategy to clarify these issues is to undertake a systematic
search for a model of minimum complexity that fits the available observations. Such an effort should start from a very simple model, perhaps
similar to a nutrient-restoring one. Complexity should then be added
only after careful analysis of the residual model-data misfits and some
educated guess about the direction of complexity enhancement.
This approach for future model improvements should consider not
only applications to the open ocean, where biogeochemical measurements are sparse and difficult to take, but also to controllable and
nipulable field experiments, e.g., in artificial enclosures, and to physiological studies on cultures in the lab. A promising example is the
application of ecosystem models and parameter optimisation methods
to mesocosm experiments [Vallino, 2000]. Mesocosms are enclosed and
generally gently mixed water volumes (typically several cubic meters in
size), either in sea-water filled tanks on land, or in large plastic bags
in the sea. They can be regarded as essentially homogeneous zerodimensional systems that, in contrast to typical liter-sized incubation
bottles, are large enough to keep boundary and enclosure effects small
for a few generation times, i.e., days. By allowing for manipulation of
environmental conditions and virtually unlimited access to observations,
combining such experiments with modelling studies via data assimilative
approaches may greatly help to better constrain our ecosystem models.
On longer time scales, the same models will also have to be validated
against open-ocean data collected within time-series programs and process studies. Physical models run in data assimilation mode may prore
543
more degrees of freedom to get closer to the observations. There are
thus two contradictory statements about the required model complexity. This may to some extent result from errors in the physics used to
drive the ecosystem models, but the problem is persistent even when
physical observations are used to provide a physical environment as
alistic as possible in a one-dimensional framework. Another possible
explanation for the apparently contradictory statements about ecosystem model complexity is that the NPZD-type models employed so far
may not have the right structures and hence are inherently inconsistent
with the yet unknown ecological rules of real marine ecosystems. It is
not clear whether more complex function-group type models would, in
this data assimilative respect, perform any better. Although their orderof-magnitude larger number in adjustable parameters should allow for
a much better fit to the data, such a fit would be of little value (and
correspond to overfitting few data points by a high-order polynomial)
for applying such models to other climate conditions unless all of the
model parameters can be constrained by observations.
A possible strategy to clarify these issues is to undertake a systematic
search for a model of minimum complexity that fits the available observations. Such an effort should start from a very simple model, perhaps
similar to a nutrient-restoring one. Complexity should then be added
only after careful analysis of the residual model-data misfits and some
educated guess about the direction of complexity enhancement.
This approach for future model improvements should consider not
only applications to the open ocean, where biogeochemical measurements are sparse and difficult to take, but also to controllable and
nipulable field experiments, e.g., in artificial enclosures, and to physiological studies on cultures in the lab. A promising example is the
application of ecosystem models and parameter optimisation methods
to mesocosm experiments [Vallino, 2000]. Mesocosms are enclosed and
generally gently mixed water volumes (typically several cubic meters in
size), either in sea-water filled tanks on land, or in large plastic bags
in the sea. They can be regarded as essentially homogeneous zerodimensional systems that, in contrast to typical liter-sized incubation
bottles, are large enough to keep boundary and enclosure effects small
for a few generation times, i.e., days. By allowing for manipulation of
environmental conditions and virtually unlimited access to observations,
combining such experiments with modelling studies via data assimilative
approaches may greatly help to better constrain our ecosystem models.
On longer time scales, the same models will also have to be validated
against open-ocean data collected within time-series programs and process studies. Physical models run in data assimilation mode may prore