Analysing airborne CO2 flux measurements in relation to spatiotemporal characteristics of drained fen meadows in the Netherlands with machine learning

Worldwide, peatlands have been transformed from carbon sinks to carbon sources due to years of intensive agriculture and livestock farming, requiring low water tables. In the Netherlands, emissions from drained organic soils mount up to around 3% of all national greenhouse gas emissions, and account for 4.6 to 7 Mt CO₂ annually. As part of the 2019 national climate agreement, the Dutch government set a specific mitigation target for these emissions of 1 Mt per year by 2030. In light of this target, the Netherlands Research Progamme on Greenhouse gas dynamics in Peatlands and organic soils (Dutch: NOBV) investigates the efficiency of proposed mitigation measures, and aims to contribute valuable scientific findings to the politically and socially sensitive debate around this issue. The focus of our study aligns with the NOBV's objective to enhance the understanding and quantification of drivers of regional emissions.

To research regional fluxes, NOBV incorporates a new approach: Eddy Covariance (EC) measurements are taken from a low-flying ultra-light aircraft. Fluxes of CO­₂, momentum, sensible and latent heat are measured, as well as meteorological variables. Weather permitting, the airborne surveys are done twice a week since 2020, over the three main fen meadow areas in the Netherlands. The crosswind, parallel flight tracks ensure that the footprints overlap, thus cover the full area. Additionally to the airborne data collection, a large EC tower network has been established with both stationary and mobile systems, encompassing 25 measurement sites.

In this study, we combine airborne and tower flux data, to make use of their different strengths: spatial heterogeneity and temporal continuity, respectively. We use footprint analysis to extract the corresponding spatial information from maps, remote sensing, and a daily soil-water information product. Using this data, we train a boosted regression tree (BRT) machine learning algorithm. Feature selection and hyperparameter tuning are applied as model optimization techniques, and subsequently Shapley values and various simulations are used to interpret the model’s outputs.

Related to the public debate and other studies on emissions from organic soils, we specifically investigate the influence of water table dynamics. A first analysis shows that during nighttime and at high incoming photosynthetically active radiation, every 10 centimeters lowering of efficient water table depth leads to 3.7 tonnes CO₂ ha^-1 yr^-1, which corresponds to current estimates. We will present these, and further results, showing what and how determines the CO₂ fluxes from drained fen meadows in the Netherlands.