|Bioindication of chemical and hydromorphological habitat characteristics with benthic macro-invertebrates based on Artificial Neural Networks|
Schleiter, I.M.; Obach, M.; Borchardt, D.; Werner, H. (2001). Bioindication of chemical and hydromorphological habitat characteristics with benthic macro-invertebrates based on Artificial Neural Networks. Aquat. Ecol. 35(2): 147-158
In: Aquatic Ecology. Springer: Dordrecht; London; Boston. ISSN 1386-2588, more
Freshwater ecology; General regression neural networks; Pollution control; Prediction; Fresh water
|Authors|| || Top |
- Schleiter, I.M.
- Obach, M.
- Borchardt, D.
- Werner, H.
This study aims to enhance the discussion about the usefulness of Artificial Neural Networks and specific input relevance detection for water quality assessment. The focus is on the development of neural modelling techniques initiating further research on predictor selection for bioindication. We tested the predictability of abiotic variables and quality indices BOD_5, conductivity, NH_3-N, NH_4-N, NO_2-N, NO_3-N, N_total, oxygen, pH-value, P_total, water temperature, chemical and morphological water quality class and saprobic index by means of benthic macro-invertebrates on 51 sampling sites of nine small streams in Central Germany. The results show that General Regression Neural Networks and modified Multi-Layer-Perceptrons can successfully be applied for modelling and predicting ecological and environmental data because of their ability to solve non-linear and multidimensional problems. Nevertheless, Linear Neural Networks have been proved suitable in some cases. Particularly, stepwise method, genetic algorithms and sensitivity analysis can be used to reduce the complexity of data sets in a reasonable way by detecting important predictors. In many cases the prediction accuracy even increases. In addition, using only the presence of species instead of their abundance provides mostly better results, simpler models and an easier collection of data. Thus, complex systems can be illustrated in easily surveyed models with low measuring and computing effort. We claim that the identification of indicator species and the assessment of complex anthropogenic impacts can be improved substantially and managed more efficiently using the neural-based approach. It is predestinated for bioindication, particularly with regard to aquatic ecosystems.