Groundwater–NEE Relationship in Dutch Peatlands Derived by Machine Learning Using Airborne and Ground-Based Eddy Covariance Data

Peatlands worldwide have been transformed from carbon sinks to carbon sources due to years of intensive agriculture requiring low water tables. In the Netherlands, carbon dioxide (CO2) emissions from drained peatlands mount up to 5.6 Mton annually and, according the Dutch climate agreement, should be reduced by 1 Mton in 2030. It is generally accepted that mitigation measures should include raising the water level, and the exact influence of water table depth has been increasingly studied in recent years. Most studies do this by comparing annual Eddy Covariance (EC) site-specific CO2 budgets to mean annual effective water table depths (WTDe). However, here we apply a different approach: we integrate measurements from 16 EC towers with EC measurements from 141 flights by a low-flying research aircraft, in an interpretable machine learning framework. We make use of the different strengths of tower and airborne data, temporal continuity and spatial heterogeneity, respectively. We apply time frequency wavelet
analysis and a footprint model to relate the measured fluxes to the underlying surface. Using spatio-temporal data, we train and optimize a boosted regression tree (BRT) machine learning algorithm and use Shapley values and various simulations to interpret the model’s outputs. We find that emissions increase with 4.6 tonnes CO2 ha-1 yr-1 for every 10 cm WTDe up to a WTDe of 0.8 meter. For more drained conditions, emissions decrease again, following an optimum-based curve. Furthermore, we find that this effect is stronger in winter than in summer and that it varies between sites. This study shows the added value of using ML with different types of instantaneous data, and holds potential for future applications.