Estimation of dissolved oxygen in riverine ecosystems: Comparison of differently optimized neural networks

Abstract

Dissolved oxygen is one of the most important water quality parameters in relation to aquatic life, and one of the most direct indicators of water pollution. The present study employs a novel methodology for the estimation of riverine dissolved oxygen using neural network models taking the spatial homogeneity of the water quality sampling sites into account. In three alternative configurations, a multivariate linear regression model, a radial basis function neural network and a general regression neural network are applied to discover which is best able to forecast dissolved oxygen effectively. Data from 13 water quality monitoring stations of the River Tisza are used, including runoff, water temperature, electric conductivity and pH. The three configurations are as follows: (i) a randomly chosen training and test set (which is then considered the reference model), (ii) a training and test set chosen in a spatially controlled way and (iii) a data set composed of two homogeneously behaving sampling sites with the addition of the site lying closest to these downstream. It was found that if the input of the linear model or the neural networks consists of a group or groups of sampling sites displaying homogeneous behavior, better performance is achieved in the estimation. Specifically, the best estimation of dissolved oxygen was achieved in the middle and lower reaches of the river, with an average of 81% and 87% of variance explained, respectively; the General Regression Neural Networks gave the best performance, in the middle reaches 85%, and 90% in the lower ones, even in the presence of a high degree of anthropogenic activity, as is the case with the River Tisza.