Predicting high-frequency nutrient dynamics in the Danube River from surrogates with sensors and machine-learning

Nutrient dynamics play an essential role in aquatic ecosystems. Despite advances in sensor technology, nutrient concentrations are difficult and expensive to monitor in-situ and in real-time. Emerging data-driven methods may provide surrogate measures for nutrient concentrations. In this work, 4-year high-frequency (15-min interval) regularly monitored variables and 2 data-driven algorithms are used to build surrogate measures for nitrate, orthophosphate, and ammonium at 2 stations in the German part of the Danube River. The variables used as input futures are dissolved oxygen (DO), temperature (Temp), conductivity (EC), pH, discharge rate (Q), and chlorophyll-a (Chl-a). Multiple linear regression (MLR) and Random Forest Regression (RF) are trained and cross-validated for the concentration predictions of nutrient constituents. Prior to training, pre-processing procedures were implemented, including removing outliers and filling missing values by linear interpolation. This work presented a thorough description of the workflow, including intermediate steps for feature engineering, feature selection, hyper-parameter optimization. The results of the 12 surrogate models (2 algorithms * 3 constituents * 2 stations) are compared. The results show that the RF algorithm can reproduce the environmental phenomena and contribute to water quality management. The RF algorithm already outperformed MLR when adding at least three predictors in this work. The five-fold CV has identified the reliable and stable prediction of the targets NO₃^--N (R² = 0.9967 and 0.9992), NH₄⁺-N (R² = 0.9861 and 0.9927), PO₄^3--P (R² = 0.9638 and 0.9643).