Determination of total phosphorus concentration in water by using visible-near-infrared spectroscopy with machine learning algorithm

Total phosphorus (TP) content is a crucial evaluation parameter for surface water quality assessment, which is one of the primary causes of eutrophication. High-accuracy, fast-speed approach for the determination of low-concentration TP in water is important. We proposed a rapid, highly sensitive, and pollution-free approach that combines spectroscopy with a machine learning algorithm we improved called synergy interval Extra-Trees regression (siETR) to determine TP concentration in water. Results show that the prediction model based on siETR can get a high coefficient of determination of prediction (Rp2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{p}^{2}$$\end{document} = 0.9444) and low root mean square error of prediction (RMSEP = 0.0731), which performs well on the prediction of TP concentration. Furthermore, the statistical analysis results further prove that the model based on siETR is superior to other models we studied both in prediction accuracy and robustness. What is more, the prediction model we established with only 140 characteristic wavelengths has the potential for the development of miniature spectral detection instruments, which is expected to achieve in situ determination of TP concentration. These results indicate that Vis–NIR spectroscopy combined with siETR is a promising approach for the determination of TP concentration in water.