Gaussian process regression for prediction of sulfate content in lakes of China

Abstract

In recent years, environmental pollution has become more and more serious, especially water pollution. In this study, the method of Gaussian process regression was used to build a prediction model for the sulphate content of lakes using several water quality variables as inputs. The sulphate content and other variable water quality data from 100 stations operated at lakes along the middle and lower reaches of the Yangtze River were used for developing the four models. The selected water quality data, consisting of water temperature, transparency, pH, dissolved oxygen conductivity, chlorophyll, total phosphorus, total nitrogen and ammonia nitrogen, were used as inputs for several different Gaussian process regression models. The experimental results showed that the Gaussian process regression model using an exponential kernel had the smallest prediction error. Its mean absolute error (MAE) of 5.0464 and root mean squared error (RMSE) of 7.269 were smaller than those of the other three Gaussian process regression models. By contrast, in the experiment, the model used in this study had a smaller error than linear regression, decision tree, support vector regression, Boosting trees, Bagging trees and other models, making it more suitable for prediction of the sulphate content in lakes. The method proposed in this paper can effectively predict the sulphate content in water, providing a new kind of auxiliary method for water detection.