Ultraviolet‐visual spectroscopy estimation of nitrate concentrations in surface waters via machine learning

Abstract

AbstractHigh‐frequency acquisition of nutrient concentrations in rivers is needed to generate nutrient loading estimates commensurate with flow and discharge data. Although the combination of field sampling and laboratory analysis is the standard approach to riverine water quality analysis, this strategy is expensive and can miss important storm‐related events. Ultraviolet‐visual (UV–Vis) spectroscopy is widely used in drinking water and wastewater systems for high‐frequency concentration estimates. However, surface waters present a unique challenge as co‐occurring constituents in environmental samples cause spectral interference at the wavelengths used to measure concentrations of dissolved nutrients. Partial least squares regression (PLSR), Lasso regression (Lasso), and stepwise multivariate linear regression (Stepwise) models can be effective predictors of nitrate concentrations using UV–Vis absorbance and are used in many available in‐situ nitrate sensors; however, the proliferation of user‐friendly open‐source machine learning (ML) algorithms offers an opportunity to use sophisticated big‐data techniques to predict nutrient concentrations in surface waters. We collected samples from four rivers across southern Ontario with a variety of nitrate concentrations, flow regimes, and interfering co‐contaminants. We demonstrated that ML applications of random forest and gradient boosting models significantly outperformed PLSR, Lasso, and Stepwise methodologies to estimate nitrate concentrations in complex environmental samples via UV–Vis absorbance. Importantly, ML applications outcompete current models at low concentrations. This new predictive methodology provides regulators and stakeholders an opportunity to establish low cost, continuous monitoring environmental programs using UV–Vis approaches.