Development of statistical regression and artificial neural network models for estimating nitrogen, phosphorus, COD, and suspended solid concentrations in eutrophic rivers using UV-Vis spectroscopy

Abstract

River water quality monitoring is crucial for understanding water dynamics and formulating policies to conserve the water environment. In situ ultraviolet-visible (UV-Vis) spectrometry holds great potential for real-time monitoring of multiple water quality parameters. However, establishing a reliable methodology to link absorption spectra to specific water quality parameters remains challenging, particularly for eutrophic rivers under various flow and water quality conditions. To address this, a framework integrating desktop and in situ UV-Vis spectrometers was developed to establish reliable conversion models. The absorption spectra obtained from a desktop spectrometer were utilized to create models for estimating nitrate-nitrogen (NO₃-N), total nitrogen (TN), chemical oxygen demand (COD), total phosphorus (TP), and suspended solids (SS). We validated these models using the absorption spectra obtained from an in situ spectrometer. Partial least squares regression (PLSR) employing selected wavelengths and principal component regression (PCR) employing all wavelengths demonstrated high accuracy in estimating NO₃-N and COD, respectively. The artificial neural network (ANN) was proved suitable for predicting TN in stream water with low NH₄-N concentration using all wavelengths. Due to the dominance of photo-responsive phosphorus species adsorbed onto suspended solids, PLSR and PCR methods utilizing all wavelengths effectively estimated TP and SS, respectively. The determination coefficients (R²) of all the calibrated models exceeded 0.6, and most of the normalized root mean square errors (NRMSEs) were within 0.4. Our approach shows excellent efficiency and potential in establishing reliable models monitoring nitrogen, phosphorus, COD, and SS simultaneously. This approach eliminates the need for time-consuming and uncertain in situ absorption spectrum measurements during model setup, which may be affected by fluctuating natural and anthropogenic environmental conditions.

Keywords: Artificial neural network; In situ UV–vis spectroscopy; Statistical regression models; Water quality monitoring; Wavelength selection.