Prediction of Stream Dissolved Oxygen Concentration in the Missouri River Basin using Feedforward Neural Network

Dissolved oxygen (DO) is necessary for the survival and growth of aquatic organisms. Accurate prediction of DO concentration is crucial to improve stream water quality and develop water resources management policies. The purpose of this study was to investigate the suitability of feedforward neural network (FFNN) models for the prediction of DO concentration in freshwater streams through an appropriate selection of dominant environmental drivers. Monthly to quarterly observational data for DO concentration and concurrent environmental variables (stream temperature, Tw; solar radiation, Rn; atmospheric pressure, Pa; pH; total phosphorus, TP; total nitrogen, TN; flowrate, Q; and specific conductivity, SC) were gathered for 13 stream monitoring stations across the Missouri River Basin during 1998-2015. First, we developed univariate (i.e., single predictor) models with the different environmental variables to identify the dominant drivers of stream DO. Multivariate FFNN models were then developed to explore the different combinations of the identified drivers and identify an optimal set to accurately predict stream DO. The univariate FFNN models indicated Tw as the major driver of stream DO, whereas Rn, SC, pH, and TP were also notable predictors. Among the developed multivariate FFNNs, the most successful and parsimonious model of stream DO was achieved by using Tw and pH as the predictors (Nash-Sutcliffe efficiency, NSE = 0.79). Further, stream DO was also successfully predicted by using Tw and TP with a slightly lower accuracy (NSE = 0.77). The findings would help obtain high quality predictions of stream DO with fewer variables for the management and restoration of ecosystem health in freshwater streams across the Missouri River Basin and beyond.