Article ID: | iaor2006156 |
Country: | Netherlands |
Volume: | 174 |
Issue: | 1/2 |
Start Page Number: | 161 |
End Page Number: | 173 |
Publication Date: | May 2004 |
Journal: | Ecological Modelling |
Authors: | Dedecker A.P., Goethals P.L.M., Gabriels W., Pauw N. De |
Keywords: | geography & environment, water, neural networks |
To meet the requirements of the EU Water Framework Directive, models are useful to predict communities in watercourses based on the abiotic characteristics of their aquatic environment. For that purpose back-propagation Artificial Neural Network (ANN) algorithms were used to induce predictive models on a dataset of the Zwalm river basin (Flanders, Belgium). This dataset consisted of 120 samples, collected over a 2-year period. Fifteen environmental variables were measured at each site, as well as the abundance of the aquatic macroinvertebrate taxa. Different neural networks were developed and optimized to obtain the best model configuration for the prediction of the habitat suitability of macroinvertebrate taxa. The best performing number of hidden layers and neurons and training algorithms have been searched for. The different options were theoretically and practically validated and assessed. The theoretical validation was based on cross-validation. For the practical validation, potential applications of the neural network models were analyzed, and the predictive performance of the models was assessed using ecological expert knowledge. The results indicate that the number of times a taxon was found in the whole river basin influences the performance measures and the architecture of the network. Based on the Cohen's kappa, it could be concluded that ANN models predicting the presence/absence of very rare taxa (e.g. Aplexa) or very common taxa (e.g. Tubificidae) were rather irrelevant, although their correctly classified instances (CCI) were high. Predicting the presence/absence of Asellidae (a moderately present taxon), the highest performances (CCI and Cohen's kappa) were found for the network model with two hidden layers each having 10 neurons. When calculation time was also taken into account, the network model with one hidden layer having 10 neurons could be preferred. Applying this network architecture, performances were only slightly worse, while calculation time was a lot shorter. One may also conclude that not all network models resulted in a relevant relation between a variable and a specific taxon. For Gammaridae for example, a rather small ANN structure gave a better idea of the impact of dissolved oxygen on its presence than a larger one. More reliable predictions and ecological interpretations for river ecosystem management would thus be possible provided the best configuration could be found.