Rasmus Rumph Frederiksen

Study regionVejrumbro, a shallow drained riparian peatland in the Nørreå River valley, central Jutland, Denmark.Study focusRiparian peatlands exhibit strong spatial and temporal heterogeneity. Consequently, controls on water table levels are not constant. We investigated how dominant controls shift through time and space using a high-resolution (5 m x 5 m) physically based model and a distance-based generalized sensitivity analysis.

The environmental and economic impacts of nitrate leaching from agricultural soils highlight the need for accurate prediction to support effective nitrogen management. This study evaluated six machine learning (ML) models—Multiple Linear Regression, Elastic Net, K-Nearest Neighbors, Decision Trees, Extra Trees, and Multi-Layer Perceptron—using 2993 field observations to predict nitrate leaching. The models incorporated crop sequence, manure and fertilizer inputs, soil, and percolation and are benchmarked against the empirical NLES5 model used in Danish nitrogen regulation. Four ML models achieved higher predictive accuracy than NLES5 when evaluated on independent test data, with the Extra Trees model achieving the best performance (R² = 0.63, RMSE = 23.3 kg N ha⁻¹), exceeding NLES5 (R² = 0.39, RMSE = 29.8 kg N ha⁻¹). Model interpretability analyses identified winter percolation, winter vegetation cover, and soil type as key drivers of nitrate leaching. The Extra Trees model was further evaluated using scenario analyses of long-term leaching trends, marginal responses to spring-applied mineral nitrogen, and spatial patterns within the Bolbro Bæk catchment. Findings highlight the potential of ML models to improve nitrate leaching predictions in ungauged areas. Future research could incorporate additional variables, including crop yield and tillage practices, to enhance model accuracy and support sustainable agricultural practices that maintain productivity while reducing nitrate leaching.

Rewetting of agricultural peatlands is widely recognised as a key climate mitigation strategy, yet its net effect remains uncertain, particularly during the early transition years. We conducted a multi-year study using continuous eddy covariance measurements to quantify carbon dioxide (CO2) and methane (CH4) fluxes across two peatland fields undergone rewetting and an adjacent shallow-drained control field in Denmark. Rewetting raised the mean annual water table level by ~7 cm relative to the control when side-by-side comparisons were possible. Our results demonstrate that rewetting can rapidly shift the ecosystem carbon balance toward net CO2 uptake, likely due to the successful establishment of productive vegetation such as Phalaris arundinacea prior to rewetting. Early vegetation development may therefore accelerate CO2 uptake compared with slower trajectories observed in more degraded peatlands. However, this benefit can be partially offset by increased CH4 emissions, particularly during wet periods, which can rival the CO2 sink strength and reduce the overall greenhouse gas mitigation potential. Even in shallow-drained control areas, modest increases in water table during wet years were sufficient to temporarily reverse net carbon loss, highlighting the sensitivity of early outcomes to hydrological conditions. Management practices, such as autumn biomass cutting, further influenced CO2 exchange by enhancing early-season uptake, though biomass removal can partially offset these gains. Collectively, our findings underscore that the net climate effect of peatland rewetting depends strongly on interannual variability, water table dynamics, vegetation establishment, and management interventions. Long-term, ecosystem-scale monitoring is thus essential to capture the full spectrum of environmental variability and optimise restoration strategies for effective climate mitigation.

High total phosphorus (TP) concentrations in freshwater such as streams, rivers, and lakes remain a global issue, causing eutrophication and poor ecological conditions. We extracted annual flow‐weighted TP concentration data ( N = 3144) from 207 monitored Danish headwater streams (catchment area < 50 km ² ) during the period 1990–2019 as the basis for developing a machine learning (ML) model for diffuse phosphorus in streams, using 22 potential predictor variables. We tested 70 different algorithms, applying an AI platform and a random split strategy with a holdout dataset (20%), a validation dataset (16%), and four training datasets (4 × 16%) in a fivefold cross‐validation procedure across a total of 207 stream stations. The ML algorithm “eXtreme Gradient Boosted trees regressor: XGBoost” was identified as the best‐performing model, based on 13 predictor variables, with a relatively high explanatory power for the training dataset ( R ²  = 0.68), the cross‐validation dataset ( R ²  = 0.71), and the holdout dataset ( R ²  = 0.66). The most important catchment characteristics of the 13 predictor variables were paved area, tile drained area, clay content in subsoil, farmed area, and TP loss from bank erosion. An external test of model transferability on a dataset, using an additional 142 stream stations ( N = 1261), revealed a somewhat lower predictive power of the final model ( R ²  = 0.41). The final model was applied to simulate annual diffuse flow‐weighted TP concentrations for 3200 unique headwater catchments (ca. 15 km ² ) and this analysis now supports the calculation of annual TP loadings to surface waters from otherwise ungauged coast‐near areas in Denmark.

Understanding and quantifying total organic nitrogen (TON) concentrations in streams and their spatial variation is essential for accurately assessing their importance for total nitrogen (TN) loadings to coastal waters and the possible sources of TON in the landscape. Total organic nitrogen constitutes almost 20% of the TN riverine loadings to Danish coastal waters. We used environmental monitoring data from 390 stations across Denmark to calculate indirectly measured annual average TON concentrations using a wide range of predictor variables. We then trained a machine learning model to predict spatially distributed average annual TON concentrations in Danish streams, achieving a mean error of 0 mg L⁻¹ and a root‐mean‐squared error of 0.20 mg L⁻¹. The mean annual predicted (measured) TON concentrations in Danish streams were 0.84 (0.70) mg L⁻¹, with a standard deviation of 0.36 (0.31) mg L⁻¹. The model is primarily driven by mean elevation and the percentages of agricultural land, tile‐drained areas, lakes and carbon‐enriched soils in the catchment. The developed model contributes to our understanding of the spatial variation in annual TON concentrations in streams at a national scale, supporting our understanding of processes driving nitrogen cycling.

Total organic nitrogen (TON) constitutes almost 20% of the total nitrogen (TN) riverine loadings to Danish coastal waters. Thus, knowledge about the TON concentrations in streams and its spatial variation is essential to accurately assess the importance of TON for TN loadings to coastal waters and thereby achieving a more precise basis for calculation of the sources of TON in catchments.We used environmental monitoring data from 390 stream stations across Denmark for the period 2018-2021to calculate indirectly measured annual and seasonal average TON concentrations (~1,500 samples) along with a wide range of predictor variables. TON samples showed a mean annual TON concentration in Danish streams amounting to 0.70 mg L-1 with a standard deviation of 0.31 mg L-1 and revealed a relatively high spatial variability.We trained a machine learning model to learn spatial and temporal patterns in our TON data set for prediction of spatially distributed annual and seasonal average TON concentrations in Danish streams in ungauged basins. Furthermore, we utilized quantile regression to estimate the uncertainty on model predictions, and we utilized quantile regression in combination with the Shapley additive explanations (SHAP) approach to investigate how the importance and influence of predictor variables vary across TON’s entire distribution.The annual TON concentration is modelled with a root-mean-squared error of 0.20 mg L-1. The new national annual average TON concentration model is largely driven by the mean elevation (negative), the percentage of agricultural land (positive), the percentage of tile drained areas (positive), and the percentage of lakes (positive).The predicted annual average TON concentrations were generally higher than the measured average annual TON concentrations, with an overall mean of 0.84 mg L-1, probably because catchments in the training data generally had higher mean elevations (DEM) than the prediction catchments as many ungauged catchments are located near the coastThe developed model and national TON maps contribute to our understanding of annual TON concentrations in streams supporting national-scale land-use and water management.

Worldwide, farming activities exert strong impacts on the amount and molecular composition of dissolved organic matter (DOM), which constitutes an important vector of organic nitrogen (ON) transport from soils to the aquatic environment (Graeber et al., 2015). However, there are major knowledge gaps on the drivers of ON loss to water courses. In Denmark, stream data from the Danish national monitoring program (NOVANA) shows that total ON currently accounts for nearly 20 % of the annual total N loading to Danish coastal waters. In a recently initiated research project ‘orgANiC’ we are investigating the loss and fate of ON forms in five smaller agricultural catchments across Denmark (Petersen et al., 2021).We are measuring dissolved ON (DON) and particulate ON as well as dissolved organic matter (DOM) and particulate organic matter (POM) in various source waters (soil water and groundwater), pathways (tile drains and surface runoff), and receiving streams using a comprehensive array of sampling technologies. In soil water we utilize suction cups taking weekly composite water samples, in groundwater we sample from near-surface (app. 1-5 m below surface) screens in boreholes using the Montejus principle, and in tile drains, surface runoff from fields and streams we are taking both grab samples and automated ISCO samples. These are activated when the hydrograph levels and hydrograph gradients exceed certain thresholds, determined from analysis of the long-term hydrograph data.We are performing both indirect (total N minus inorganic N) and direct analysis of DON (size exclusion chromatography) on water samples from the different hydrological compartments. The loss of particulate ON (PON) is also monitored in tile drainage water, surface runoff and streams as these three hydrological paths are believed to be of increasing importance with the observed increase in extreme weather conditions. In the presentation we will share our current insights into the challenges of indirect DON measurements across different hydrological pathways by comparing it with direct measurements of DON and PON. We will also demonstrate how the concentrations and composition of ON fractions vary across the agricultural catchments under investigation as they represent different soil types, climate conditions and agricultural management (crops, fertilization, etc.). ReferencesGraeber, D., I. G. Boëchat, F. Encina-Montoya, and others. 2015. Global effects of agriculture on fluvial dissolved organic matter. Scientific Reports 5: 16328. doi:10.1038/srep16328.Petersen, RJ, Blicher-Mathiesen, G, Rolighed, J, Andersen, HE & Kronvang, B 2021, 'Three decades of regulation of agricultural nitrogen losses: Experiences from the Danish Agricultural Monitoring Program', Science of the total Environment 787: 147619. https://doi.org/10.1016/j.scitotenv.2021.147619

Temporal drain flow dynamics and understanding of their underlying controlling factors are important for water resource management in tile-drained agricultural areas. The use of physics-based water flow models to understand tile drained systems is common. These models are complex, with large parameter sets and require high computational effort. The primary goal of this study was to examine whether simpler, more efficient machine learning (ML) models can provide acceptable solutions. The specific aim of our study was to assess the potential of ML tools for predicting drain flow time series in multiple catchments subject to a range of climatic and landscape conditions. The investigation is based on unique data containing time series of daily drain flow in multiple field scale drain sites in Denmark. The data include: climate (precipitation, potential evapotranspiration, temperature); geological properties (clay fraction, first sand layer thickness, first clay layer thickness); and topographical indexes (curvature, Topographical wetness indexes, Topographical position index, elevation). Both static and dynamic variables are used in the prediction of drain flows. The ML algorithm extreme gradient boosting (XGBoost) and convolutional neural network (CNN) were examined, and the results were compared with a physics-based distributed model (MIKE-SHE). The results show that XGBoost performs similarly to the physics-based MIKE-SHE models, and both outperform CNN. Both ML models required significantly less effort to build, train, and run than MIKE-SHE. In addition, the ML models support efficient feature importance analysis. This showed that climatic variables were important for CNN models and XGBoost. The results support the use of ML models for hydrologic applications with sufficient data for training. Further, the insights offered by the feature importance analysis may support further data collection and developments of physics-based models when existing data are insufficient to support ML approaches.