You are on page 1of 5

August 24, 2011 11:45 World Scientific Book - 9.75in x 6.

5in Main

Chapter 12

Modeling Philosophy
A Guide to Modeling Coastal Morphology Downloaded from www.worldscientific.com

12.1 Virtual reality or realistic analogue?


by LA TROBE UNIVERSITY on 01/01/16. For personal use only.

As we have seen in the previous chapters, process-based morphodynamic modeling


of coastal areas has taken off in recent years, due to robust numerical schemes, in-
creased computer power and clever tricks to speed up the simulations. It is allowing
engineers and scientists to use models as numerical laboratories of which the con-
ceptual handling is similar to that in physical scale models: create bathymetries,
fill with sediment (‘define initial sediment thickness’), turn on pumps (‘discharge
boundaries’), create weirs (‘clamped water level boundaries’) and place wave pad-
dles (‘incident spectral wave conditions’) to create boundary conditions and analyze
what happens in terms of evolving bed patterns and sedimentation/erosion rates.
Now that we have these numerical labs, an important question is how best to use
them to solve real-life problems. In this section we will discuss pros and cons of two
fundamentally different approaches.

Virtual reality... The first, most common method applied today is to try and
create a ‘virtual reality’: we reproduce in the best possible detail the geometry,
sedimentology, bathymetry, 3D flow, wave and sediment processes and simulate the
evolution of the sea or river bed and associated processes under various future sce-
narios, in order to find out likely effects of something we intend to do (deepen a
channel, build a new port, reclaim land...). The success of this approach depends
critically on our ability a) to prove that our model accurately represents the mor-
phological processes going on at present and b) to prove that it is still valid when
the situation changes drastically in the future. The method to validate a) is usually
to carry out a hindcast over a recent period where our model results are calibrated
against the observed hydrodynamics, sediment concentration patterns, sedimenta-
tion/erosion patterns and dredging quantities. Here we encounter the problem that,
no matter how good our model is, it is not reality and it may not accept the initial
conditions we offer it as the near-equilibrium conditions that they are. This will
lead our model to quickly try to adjust to a more fitting bathymetry. This will
show up in the results as sedimentation/erosion patterns that are different from the

255
August 24, 2011 11:45 World Scientific Book - 9.75in x 6.5in Main

256 Coastal Morphodynamic Modeling Guide

observations, which can pose serious problems, for example, if one investigates ways
to optimize dredging and the model predicts erosion where most of the dredging
takes place. Unless one can identify and fix the root cause for such a discrepancy,
‘tuning’ the model may seemingly improve the hindcast but actually worsen its
longer-term behaviour, as illustrated in Figure 12.1. Of course, this pessimistic
scenario need not always be true, but there is a real danger here that we need to
be aware of. Taking this route we can only slowly improve the confidence in our
models by making sure and proving that they contain all the relevant physics and
by assessing them against case after case with as little adjustments as possible.
A Guide to Modeling Coastal Morphology Downloaded from www.worldscientific.com
by LA TROBE UNIVERSITY on 01/01/16. For personal use only.

Fig. 12.1 Pessimistic scenario for effect of calibration

...or realistic analogue? In the second approach we use our models differently:
we do not try to replicate the situation exactly but we use the model as an analogue
to reality, with which we can investigate processes and effects in relative isolation.
Instead of asking the question: ‘where will this particular channel be located next
year?’ we may ask: ‘why is there a channel here, with these particular dimen-
sions?’ We may investigate which hydrodynamic forcing creates and maintains it,
in how far it is self-organized or forced by geometric constraints, how important is
the geological setting, by carrying out numerical experiments where we vary such
parameters. Given some necessary simplifications such experiments can be run over
very long time periods, allowing us to study the equilibrium behaviour of the model
for specific situations (delta formation, channel structure in tidal inlets and estuar-
ies). Once a near-equilibrium situation is established, various integral parameters
of the simulated pattern can be compared with empirical relationships. While such
studies offer valuable insight into what causes different bed patterns, there is a
practical application, too. Once an ‘establishment scenario’ has been set up and a
August 24, 2011 11:45 World Scientific Book - 9.75in x 6.5in Main

Modeling Philosophy 257

situation has been created that shows realistic behaviour and patterns, this provides
a good starting point for sensitivity studies investigating in principle the effects of,
say, sea level rise, land reclamations, dredging or nourishment strategies etc.
Our aim in discussing these two approaches is not to select one or the other;
rather, we believe that they complement each other. The detailed ‘virtual reality’
models may gain much from insight gained in the ‘realistic analogue’ models and can
gain public acceptance when it is shown that the physics contained in them leads
to realistic long-term results, whereas the accuracy of the more schematic models
obviously may profit from improved process descriptions validated at detailed scale.
A Guide to Modeling Coastal Morphology Downloaded from www.worldscientific.com

12.2 Process-based or data-driven?


by LA TROBE UNIVERSITY on 01/01/16. For personal use only.

This book has been mostly about process-based modeling. We’ve seen that this kind
of modeling can explain a lot of morphological processes that we see, but this does
not mean that we can always predict what we want to. If we take the example of
predicting the behavior of longshore bars as we discussed in Chapter 7, we’ve shown
that we can hindcast the evolution of the bars, with both onshore and offshore mo-
tions, in a realistic way. However, predicting the bar behavior into the future is
another thing; it suffers from the fact that the wave climate is not deterministic
and from the build up of errors over the months and years. To represent the longer-
term behavior we need to carefully select the most relevant processes, be careful to
maintain only the necessary adjustable coefficients and to carry out an extensive
calibration. By this time, you could argue that our model has become to a large ex-
tent data-driven, so why not directly use a data-driven approach such as a (linear or
nonlinear) neural network? [Pape (2010)] has compared process-based (UNIBEST)
model results of long-term evolution of longshore bars with neural networks and
concludes that the latter outperform the process-based model in predicting param-
eters such as the bar location. His claim is that this is due to the accumulation of
errors in process-based models, but it may well be that it is just easier to train a
neural network model than a process-based model. On the other hand, once suf-
ficiently trained for longer-term evolution, process-based modeling can be used to
predict things that are not in the training dataset, such as nourishments and coastal
structures, and can still be applied, although with more uncertainty, in data-scarce
environments, especially when we use the model to predict the overall behavior and
the relative effects of engineering measures, rather than absolute locations of the
bars.

12.3 Top-down or bottom-up?

The distinction between top-down approaches (e.g. [Stive and Wang (2003)], [Plant
et al. (1999)] and bottom-up approaches is much less sharp than it seems in the
August 24, 2011 11:45 World Scientific Book - 9.75in x 6.5in Main

258 Coastal Morphodynamic Modeling Guide

kind of applications like predicting bar behavior or the evolution of tidal systems.
A real bottom-up approach would mean to calibrate a model against data over a
short period and then extrapolating the resulting model over years or centuries.
This obviously leads to a big accumulation of errors and poor results on the long
term, not only quantitatively but also qualitatively. But we can do better than
that: if we examine the model behavior on longer time scales by doing many runs,
we slowly find out which processes and parameters govern the longer-term trends
and how to set them. In other words, we then also work from the top (the desired
end result) down (to the processes and parameters needed to get there). Important
success factors in this is that the model must be robust enough to survive all these
A Guide to Modeling Coastal Morphology Downloaded from www.worldscientific.com

runs and that the processes it contains allow enough richness of behavior.
by LA TROBE UNIVERSITY on 01/01/16. For personal use only.

12.4 More physics, better model?

Though it sounds reasonable that adding more physics will improve the ability to
represent certain processes, the overall model performance does not always have to
improve. The problem is that our representations of physical processes, especially
around breaking waves and sediment transport, are not exact and every process
we add will come with at least one extra uncertain coefficient. We may then end
up with lots of adjustable coefficients, for which there is no clear guidance and for
which it is unclear how they influence the end result. A simpler model with fewer
coefficients that have a predictable effect on the outcome will then be preferable.
A nice example is given by [Ruessink et al. (2007)] who, based on a very extensive
sensitivity analysis, threw out a lot of processes in a profile model and ended up
with a much more manageable set of processes and coefficients, with which they
could hindcast profile evolution for different sites with considerable skill. Of course
we should keep in mind the Einstein Principle, which says that ”a scientific theory
should be as simple as possible, but no simpler”. The same holds for models.

12.5 How to judge model skill?

It is perhaps a sign of the growing maturity of morphological models that quan-


tifying the skill of morphological simulations is becoming a serious issue. In the
early days of intercomparing models in EU projects this was carefully avoided, but
especially during the Coast3D project important steps were taken to develop useful
metrics for our models in [Sutherland et al. (2004a)], which were then tested for a
complex field site in [Sutherland et al. (2004b)]. For hydrodynamic data, usually
time series of values, it is not so hard to come up with meaningful relative measures
of error. Evaluating the skill in predicting bottom changes is harder. Often we see
encouraging visual agreement but if we try to quantify this it turns out to be not
so good. A notorious skill measure is the Brier Skill Score (BSS), which indicates if
August 24, 2011 11:45 World Scientific Book - 9.75in x 6.5in Main

Modeling Philosophy 259

your model does better than just stating that nothing will change. Negative scores
indicate that you are worse than that, a perfect score is 1. In many practical cases
it is very difficult to come up with skill socres above 0, let alone to get close to
1. This should not discourage us from using this skill score; first of all, it will still
tell us if model improvements actually improve results and secondly, it is no use
pretending the models are better than they are. In the meantime, a bad skill score
for absolute bottom changes does not have to mean the model is all bad; a slight
shift in channel location can kill a BSS score. Then again, this shift might be just
what the client was interested in...
A Guide to Modeling Coastal Morphology Downloaded from www.worldscientific.com

12.6 Absolute vs. relative skill

Finally we come back to the question: what do we use our model for? We spend a
by LA TROBE UNIVERSITY on 01/01/16. For personal use only.

lot of time hindcasting observed morphology changes, with varying success, but is
predicting the past what we are after? These simulations mainly serve to establish
whether or not our models behave realistically, with a variability in the right order
of magnitude and with natural features generated or preserved in roughly the right
shape. Once this is the case, we can start feeling somewhat confident in using the
model for evaluating scenarios, ’what if?’ questions. If the model does not look like
reality, it is unlikely that it can predict even relative effects correctly, although this
is often suggested. If it does, it is still no guarantee that the effect of a particular
scenario we’re interested in will be modeled correctly; we will need a collection of
case studies on real executed projects to be able to make that judgement. That
way we can slowly make progress towards having useful morpological models with
known skills.

You might also like