Value and limitations of machine learning in high-frequency nutrient data for gap-filling, forecasting, and transport process interpretation.

High-frequency monitoring of water quality in catchments brings along the challenge of post-processing large amounts of data. Moreover, monitoring stations are often remote and technical issues resulting in data gaps are common. Machine learning algorithms can be applied to fill these gaps, and to a certain extent, for predictions and interpretation. The objectives of this study were (1) to evaluate six different machine learning models for gap-filling in a high-frequency nitrate and total phosphorus concentration time series, (2) to showcase the potential added value (and limitations) of machine learning to interpret underlying processes, and (3) to study the limits of machine learning algorithms for predictions outside the training period. We used a 4-year high-frequency dataset from a ditch draining one intensive dairy farm in the east of The Netherlands. Continuous time series of precipitation, evapotranspiration, groundwater levels, discharge, turbidity, and nitrate or total phosphorus were used as predictors for total phosphorus and nitrate concentrations respectively. Our results showed that the random forest algorithm had the best performance to fill in data-gaps, with R2 higher than 0.92 and short computation times. The feature importance helped understanding the changes in transport processes linked to water conservation measures and rain variability. Applying the machine learning model outside the training period resulted in a low performance, largely due to system changes (manure surplus and water conservation) which were not included as predictors. This study offers a valuable and novel example of how to use and interpret machine learning models for post-processing high-frequency water quality data.

Transport-limited kinetics of phosphate retention on iron-coated sand and practical implications.

Iron-coated sand (ICS) is a by-product from drinking water treatment made of sand coated with ferric iron (hydr)oxides. It is considered a suitable material for large-scale measures for phosphate removal from natural and agricultural waters to prevent eutrophication. Previous studies demonstrated that the residence time of water must be very long to reach equilibrium partitioning between phosphate and ICS but specifics for application are missing. First, SEM-EDX images were used to support the conceptual assumption that P adsorption inside the coating is a transport-limited process. Second, a conceptual model of phosphate adsorption was proposed considering two types of sites: one type with fast adsorption kinetics and reaching equilibrium with the percolating solution, and another type for which adsorption is also reversible but described by pseudo-first-order kinetics. The latter is conceived to account for transport-limited adsorption in the interior of the coating while the former fraction of sites is assumed to be easily accessible and located close to the grain surface. Third, the kinetics of phosphate adsorption on ICS were quantitatively determined to describe and predict phosphate retention in filters under various flow conditions. The model was calibrated and validated with long-term column experiments, which lasted for 3500 h to approach equilibrium on the slowly reacting sites. The model reproduced the outflowing phosphate concentrations: the pronounced increase after a few pore volumes and the slow increase over the remaining part of the experiment. The parameterized model was also able to predict the time evolution of phosphate concentrations in the outflow of column experiments with different flow velocities, flow interruption, and in desorption experiments. The equilibrium partition coefficient for the experimental conditions was identified as 28.1 L/g-Fe at pH 6.8 and a phosphate concentration of 1.7 mg-P / L. The optimized first-order mass transfer coefficient for the slow adsorption process was 1.56 10-4 h-1, implying that the slow adsorption process has a time scale of several months. However, based on the parameterized model, the slow adsorption process accounted for 95.5% of the equilibrium adsorption capacity, emphasizing the potential relevance of this process for practical applications. The implications for the design, operation, and lifespan of ICS filters are exemplarily illustrated for different scenarios.

Phosphorus adsorption on iron-coated sand under reducing conditions.

Mitigation measures are needed to prevent large loads of phosphate originating in agriculture from reaching surface waters. Iron-coated sand (ICS) is a residual product from drinking water production. It has a high phosphate adsorption capacity and can be placed around tile drains, taking no extra space, which increases the farmers' acceptance. The main concern regarding the use of ICS filters below groundwater level is that limited oxygen supply and high organic matter concentrations may lead to the reduction and dissolution of iron (hydr)oxides present and the release of previously adsorbed phosphate. This study aimed to investigate phosphate adsorption on ICS at the onset of iron reduction. First, we investigated whether simultaneous metal reduction and phosphate adsorption were relevant at two field sites in the Netherlands that use ICS filters around tile drains. Second, the onset of microbially mediated reduction of ICS in drainage water was mimicked in complementary laboratory microcosm experiments by varying the intensity of reduction through controlling the oxygen availability and the concentration of degradable organic matter. After 3 yr, ICS filters in the field removed phosphorus under low redox conditions. Over 45 d, the microbial reduction of manganese and iron oxides did not lead to phosphate release, confirming field observations. Electron microscopy and X-ray absorption spectroscopy did not evince systematic structural or compositional changes; only under strongly reducing conditions did iron sulfides form in small percentages in the outer layer of the iron coating. Our results suggest that detrimental effects only become relevant after long periods of operation.