Prediction of water quality based on SVR by fluorescence excitation-emission matrix and UV-Vis absorption spectrum

As a result of economic development, the pollution of freshwater resources in urban areas of China is becoming more and more serious. Therefore, it is urgent and necessary to develop a real-time monitoring method for the water quality of urban streams and rivers. In this study, a novel method (CFFA) Combined by peak-picking method, Fluorescence spectral indexes, Fluorescence regional integration, and Absorption spectral indexes were designed to extract wide-ranging information from the combination of the excitation-emission matrix (EEM) and absorption spectrum (Abs) of water samples. More than 600 freshwater samples were collected at 180 sections of 60 rivers in the Yangzhou urban region from April 2018 and May 2019. The CFFA inputs form was applied to establish the prediction models of water quality indexes (COD_Cr, COD_Mn, NH₃-N, TP, TN, and BOD₅) based on ε-Support Vector Regression (ε-SVR). To examine the performance of the prediction models, contrastive analysis among CFFA and the other three input models was carried out. Results show that CFFA input models have shorter modeling time, lower RMSE and MAPE, and higher R² in both training and testing sets, and each constituent part of CFFA is important to the precise prediction on the basis of the ablating analysis. Our study highlights that SVR models with the CFFA input trained by numerous and various water samples could effectively predict multiple indexes for real-time water quality monitoring.