104
P. Goethals· A. Dedecker· W. Gabriels . N. De Pauw
The aim of this study was to determine the most efficient restoration option to
obtain a stable biological ecosystem meeting the minimal river water quality
standards for Flanders, such as 'the Belgian Biotic Index (BB!) equal or higher
than seven'. In Tab. 6.3 only predictions for the best restoration option are
summarized, based on aselection made from a set of simulations with artificial
neural network models (Dedecker 2001). The results indicate that after river
restoration, some macroinvertebrate taxa, indicative for a good water quality and
that are currently not present, will colonize the site again. Also the predicted BBI
changes from a moderate to a good quality and illustrates that the basic water
quality standards for Flanders are met under the mitigated conditions.
6.4
Discussion
GeneraHy the classification trees performed weH to predict the macroinvertebrate
taxa, based on the fifteen input variables. This method does not merely generate
results with a low prediction error, but also allows the user to identify associations
and general trends in the data (as illustrated by the Gammaridae model), making it
more interesting than complete black-box techniques. One may conclude that
classification trees are interesting grey-box prediction techniques, aHowing the
user to combine a smaH prediction error with getting some information on general
trends in the data. This methodology can thus be used to determine the ecological
requirements of organisms that are not sufficiently weH understood (Dzeroski et
al. 1997). Probably the results can be improved by providing other valuable
inputs. Experiments with different sets of input variables did not only result in an
altered prediction error, but also the complexity of the trees as weH as the relative
importance of some 'general' trends seemed to be affected. Further research is
therefore necessary to get insight in the impact of different input variable sets on
the prediction qualities of decision trees. D'heygere et al. (2001) illustrated the
convenience of genetic algorithms to automatically select input variables sets in
this context. Also models for very common or very rare species need to be
optimised. The trees of these taxa are very limited and have in many cases no
added value to black box models and probabilistic guesses. Boosting, bagging and
meta-cost algorithms (Witten and Frank 2000) can have an interesting added value
for this, although in many cases the transparency and robustness as weH as their
ecological validity of the induced mies is often affected by these techniques.
Although mle-induction by classification trees generates in general robust
models with a high predictive reliability, one has to be aware of the static
characteristics of this type of models. Therefore, the simulations still need to be
checked by ecological experts that can deli ver knowledge that is often not
included in the database used for the model induction. This was illustrated by the
case study on the weirs-removal in the Zwalm river basin. To increase the model
feasibility with regard to simulations for river restoration management, spatialtemporal expert-mies will have to be included (such as migration kinetics of the
P. Goethals· A. Dedecker· W. Gabriels . N. De Pauw
The aim of this study was to determine the most efficient restoration option to
obtain a stable biological ecosystem meeting the minimal river water quality
standards for Flanders, such as 'the Belgian Biotic Index (BB!) equal or higher
than seven'. In Tab. 6.3 only predictions for the best restoration option are
summarized, based on aselection made from a set of simulations with artificial
neural network models (Dedecker 2001). The results indicate that after river
restoration, some macroinvertebrate taxa, indicative for a good water quality and
that are currently not present, will colonize the site again. Also the predicted BBI
changes from a moderate to a good quality and illustrates that the basic water
quality standards for Flanders are met under the mitigated conditions.
6.4
Discussion
GeneraHy the classification trees performed weH to predict the macroinvertebrate
taxa, based on the fifteen input variables. This method does not merely generate
results with a low prediction error, but also allows the user to identify associations
and general trends in the data (as illustrated by the Gammaridae model), making it
more interesting than complete black-box techniques. One may conclude that
classification trees are interesting grey-box prediction techniques, aHowing the
user to combine a smaH prediction error with getting some information on general
trends in the data. This methodology can thus be used to determine the ecological
requirements of organisms that are not sufficiently weH understood (Dzeroski et
al. 1997). Probably the results can be improved by providing other valuable
inputs. Experiments with different sets of input variables did not only result in an
altered prediction error, but also the complexity of the trees as weH as the relative
importance of some 'general' trends seemed to be affected. Further research is
therefore necessary to get insight in the impact of different input variable sets on
the prediction qualities of decision trees. D'heygere et al. (2001) illustrated the
convenience of genetic algorithms to automatically select input variables sets in
this context. Also models for very common or very rare species need to be
optimised. The trees of these taxa are very limited and have in many cases no
added value to black box models and probabilistic guesses. Boosting, bagging and
meta-cost algorithms (Witten and Frank 2000) can have an interesting added value
for this, although in many cases the transparency and robustness as weH as their
ecological validity of the induced mies is often affected by these techniques.
Although mle-induction by classification trees generates in general robust
models with a high predictive reliability, one has to be aware of the static
characteristics of this type of models. Therefore, the simulations still need to be
checked by ecological experts that can deli ver knowledge that is often not
included in the database used for the model induction. This was illustrated by the
case study on the weirs-removal in the Zwalm river basin. To increase the model
feasibility with regard to simulations for river restoration management, spatialtemporal expert-mies will have to be included (such as migration kinetics of the