Keywords
macroinvertebrates, artificial neural networks, modelling, water quality
Start Date
1-7-2004 12:00 AM
Abstract
Back-propagation Artificial Neural Networks (ANN) were tested with the aim of modelling the occurrence of benthic macroinvertebrate families in a south Brazilian river. The dataset, consisting of 67 sets of observations of macroinvertebrate abundance (families Hydrobiidae, Tubificidae, Chironomidae, Baetidae and Leptophlebiidae) and water quality variables (pH, temperature, dissolved oxygen, biochemical oxygen demand, nitrate, phosphate, total solids, turbidity and fecal coliforms), was collected at eleven sampling sites in the Sinos River Basin during 1991-1993. Five different ANN architectures, with one hidden layer and 2, 5, 10, 20 and 25 neurons were tested. The ANN models were trained using the gradient descendent and Levenberg-Marquardt (LM) algorithms, with different combinations of sigmoid transfer functions (log-log, tan-log, tan-tan). The percentage of success and the correlation coefficient were used to choose the best network architecture for each taxon. The networks with the LM algorithm provided the best predictions of macroinvertebrate family occurrence, independent of the family’s frequency. The same network architecture did not always reproduce all the relationships between the taxon occurrence and the environmental variables. The best model, based on a high correlation coefficient among real and predicted data and a high percentage of successes, was the ANN for a very common taxon (Hydrobiidae).
Benthic Macroinvertebrates Modelling Using Artificial Neural Networks (ANN): Case Study of a Subtropical Brazilian River
Back-propagation Artificial Neural Networks (ANN) were tested with the aim of modelling the occurrence of benthic macroinvertebrate families in a south Brazilian river. The dataset, consisting of 67 sets of observations of macroinvertebrate abundance (families Hydrobiidae, Tubificidae, Chironomidae, Baetidae and Leptophlebiidae) and water quality variables (pH, temperature, dissolved oxygen, biochemical oxygen demand, nitrate, phosphate, total solids, turbidity and fecal coliforms), was collected at eleven sampling sites in the Sinos River Basin during 1991-1993. Five different ANN architectures, with one hidden layer and 2, 5, 10, 20 and 25 neurons were tested. The ANN models were trained using the gradient descendent and Levenberg-Marquardt (LM) algorithms, with different combinations of sigmoid transfer functions (log-log, tan-log, tan-tan). The percentage of success and the correlation coefficient were used to choose the best network architecture for each taxon. The networks with the LM algorithm provided the best predictions of macroinvertebrate family occurrence, independent of the family’s frequency. The same network architecture did not always reproduce all the relationships between the taxon occurrence and the environmental variables. The best model, based on a high correlation coefficient among real and predicted data and a high percentage of successes, was the ANN for a very common taxon (Hydrobiidae).
