Unravelling and predicting ecosystem behaviours of Lake Soyang (South Korea) in response to seasonality and management by means of artificial neural networks

摘要

Artificial neural networks prove to be powerful tools for forecasting short-term changes in phytoplankton abundances and unravelling ecological complexity of limnological time series data. Applications of recurrent supervised ANN to time-series data from 1988 to 2000 of Lake Soyang resulted in models for 7-days-ahead forecasting of Chl-a and abundances of Anabaena and Asterionella populations. The predictions of Chl-a proved to be reasonably valid for three years with highly different fish farming intensities. The predictions of abundances of Anabaena and Asterionella proved only partially to be valid for two years with highly different fish farming intensities. Sensitivity analysis by means of recurrent supervised ANN trained for Anabaena and Asterionella revealed their relationships with water temperature and pH changes in accordance with the theory. Nonsuper-vised ANN based ordination and clustering of time-series data of Anabaena, Microcystis and Asterionel-la abundances from 1988 to 2000 of Lake Soyang according to temperature and pH ranges resulted in qualitative patterns confirming favouring water quality conditions for the three algal populations. Ordination and clustering of time-series data of Chl-a for four periods with different intensities of fish farming in Lake Soyang demonstrated the capacity of the method to assess ecological long-term effects of changing environmental conditions with respect to seasons.