Artificial neural nets and abundance prediction of aquatic insects in small streams

Abstract

Abundance prediction of aquatic insects (Ephemeroptera, Plecoptera, Trichoptera = EPT) based on environmental variables (precipitation, discharge, temperature) and abundance of the parent generation with Artificial Neural Nets (ANN) was carried out successfully. A general model for all species does not exist. Easy to understand models for individual species were restricted to stream sections with a characteristic set of variables. The amount of zero-values in the data did not affect the models. Transfer of one model to other stream sections resulted in a decrease of the determination coefficient B. Sufficient models for populations that have larvae in the stream all the year round required more information than for species with a diapause. All scaling options used decreased prediction quality. Long term mean values of variables and the deviation of actual from long term data were the best predictors, indicating a successful temporal link between seasonal variables and univoltine life cycles of most species tested. Prediction of monthly emergence in individual years was adequate with determination coefficients > 0.8 for five, and < 0.5 for only two out of ten years.