Chapter 6 . Stream Assessment
105
organisms in water, land and air), as weIl as annual measurement campaigns that
can improve the database with regard to its information contents.
The performance of the ANN models is in general better than simple
probabilistic predictions and rather similar to classification tree models. Also
Walley and Fontama (1998), Schleiter et al. (1999) and Gabriels (2000) came to
similar conclusions for predicting macroinvertebrates based on a limited set of
environmental characteristics. ANN models for common and rare taxa have the
highest reliability (expressed as correctly classified instances or CCI), while for
moderately frequent taxa the prediction is lower, but relatively much better than
the probabilistic guesses as was also reflected in the classification tree models.
Sensitivity analyses allowed to study the impact of the input variables on the
presence or absence of macroinvertebrate taxa. The impact of flow velocity on the
absence and presence of Gammaridae is confirmed by earlier observations (De
Pauw and Vannevel 1993) and also by the rules induced via the classification tree
models. Many other relations were detected and were in most cases confirmed
with related ecological research results, when this information was available. This
also indicates that these models in many cases work in an ecological meaningful
manner. In this way ANN models allow to determine the major variables that
affect the ecosystem quality and should be taken under direct consideration in the
river ecosystem management. Further research is necessary to determine the
optimal neural network configuration. Walley and Fontama (2000) used an ANN
with two hidden layers with six nodes each for similar simulations. Also the
impact of the applied training algorithms as weIl as the risk of overtraining the
network should be further analysed to obtain reliable and meaningful predictions
in the long run.
Several case studies related to restoration (e.g. Bettelhovebeek) and
environmental impact assessment proved the interesting added value of the ANN
ecosystem models developed for river management.
Acknowledgements
The authors like to thank the Scientific Research Foundation of Flanders (FWOFlanders) for its financial support (project 3GOl.02.97). Special thanks to Nico
Raes for his support with regard to the data collection and to Saso Dzeroski for his
help on the classification tree induction.
References
AMINAL (1999) Control of sediment transport in unnavigable watercourses as part of
integrated water management: Zwalm river basin project. AMINAL, Ohent (in Dutch)
Barros LC, Bassanezi RC, Tonelli PA (2000) Fuzzy modelling in population dynamics.
Ecological ModelJing, 128, 27-33
Breimann L, Friedman JH, Olshen RA, Stone CJ (1984) Classification and regression trees.
Pacific Orove,Wadsworth
105
organisms in water, land and air), as weIl as annual measurement campaigns that
can improve the database with regard to its information contents.
The performance of the ANN models is in general better than simple
probabilistic predictions and rather similar to classification tree models. Also
Walley and Fontama (1998), Schleiter et al. (1999) and Gabriels (2000) came to
similar conclusions for predicting macroinvertebrates based on a limited set of
environmental characteristics. ANN models for common and rare taxa have the
highest reliability (expressed as correctly classified instances or CCI), while for
moderately frequent taxa the prediction is lower, but relatively much better than
the probabilistic guesses as was also reflected in the classification tree models.
Sensitivity analyses allowed to study the impact of the input variables on the
presence or absence of macroinvertebrate taxa. The impact of flow velocity on the
absence and presence of Gammaridae is confirmed by earlier observations (De
Pauw and Vannevel 1993) and also by the rules induced via the classification tree
models. Many other relations were detected and were in most cases confirmed
with related ecological research results, when this information was available. This
also indicates that these models in many cases work in an ecological meaningful
manner. In this way ANN models allow to determine the major variables that
affect the ecosystem quality and should be taken under direct consideration in the
river ecosystem management. Further research is necessary to determine the
optimal neural network configuration. Walley and Fontama (2000) used an ANN
with two hidden layers with six nodes each for similar simulations. Also the
impact of the applied training algorithms as weIl as the risk of overtraining the
network should be further analysed to obtain reliable and meaningful predictions
in the long run.
Several case studies related to restoration (e.g. Bettelhovebeek) and
environmental impact assessment proved the interesting added value of the ANN
ecosystem models developed for river management.
Acknowledgements
The authors like to thank the Scientific Research Foundation of Flanders (FWOFlanders) for its financial support (project 3GOl.02.97). Special thanks to Nico
Raes for his support with regard to the data collection and to Saso Dzeroski for his
help on the classification tree induction.
References
AMINAL (1999) Control of sediment transport in unnavigable watercourses as part of
integrated water management: Zwalm river basin project. AMINAL, Ohent (in Dutch)
Barros LC, Bassanezi RC, Tonelli PA (2000) Fuzzy modelling in population dynamics.
Ecological ModelJing, 128, 27-33
Breimann L, Friedman JH, Olshen RA, Stone CJ (1984) Classification and regression trees.
Pacific Orove,Wadsworth