Prediction Machine Learning Methods for Dissolved Oxygen Value of the Sakarya Basin in Turkey

Abstract

In this study, deep learning (DL), support vector machine regression (SVMR), Gaussian process regression (GPR), and artificial neural networks (ANNs) models were used for dissolved oxygen estimation. The results of these four methods were compared. Monthly water quality at 20 stations between 1995 and 2014 in Sakarya Basin Application data is used. The involved parameters, including temperature, electrical conductivity, pH, biological oxygen demand, and time in months, are used as inputs to the models. The following measures were used to evaluate the accuracy of the results of the four methods (a) the error statistics of mean absolute error (MAE), (b) root mean square error (RMSE), (c) determination coefficient (R2), and (d) Nash–Sutcliffe efficiency coefficient (NSE). Assessment results showed that the error criteria of the four methods are close to each other. But the MAE (1.10) and RMSE (1.45) values of the ANN approach are lower than the other methods. At the same time, Taylor and Violin diagrams were used to compare the results. The results indicated that close values were obtained from the model and the observations. Therefore, it is concluded that the ANN approach performed better than other approaches in estimating the dissolved oxygen values. Although DL, SVMR, and GPR methods are generally preferred because of their superior predictive power in short datasets, they did not show the expected performance in this study. Additionally, it was determined that the tested measured data by the Kruskal–Wallis test (KW) came from the same distribution. Consequently, we could conclude that the efficiency of the methods recommended in the comparison has been proven with KW. © 2023 Elsevier B.V., All rights reserved.