River phytoplankton prediction model by Artificial Neural Network: Model performance and selection of input variables to predict time-series phytoplankton proliferations in a regulated river system

In this study, a comparison between statistical regression model and Artificial Neural Network (ANN) is given on the effectiveness of ecological model of phytoplankton dynamics in a regulated river. From the results of the study, the effectiveness of ANN over statistical method was proposed. Also feasible direction of increasing ANN models' performance was provided. A hypertrophic river data was used to develop prediction models (chlorophyll a (chl. a) 41.7 ± 56.8 μg L− 1; n = 406). Higher time-series predictability was found from the ANN model. Failure of statistical methods would be due to the complex nature of ecological data in the regulated river ecosystems. Reduction of ANN model size by decreasing the number of input variables according to the sensitivity analysis did not have effectiveness with respect to the predictability on testing data set (RMSE of the ANN with all 27 variables, 25.7; 47.9 from using 2 highly sensitive variables; 42.9 from using 5 sensitive variables; 33.1 from using 15 variables). Even though the ANN model presented high performance in prediction accuracy, more efficient methods of selecting feasible input information are strongly requested for the prediction of freshwater ecological dynamics.