Deciphering eddy-covariance CO₂ flux patterns in Dutch peatlands, from machine-learning to physics-based deep-learning

Peat soil degradation in the Netherlands contributes an estimated 4.6–7 Mt CO₂ annually, accounting for approximately 3% of national greenhouse gas emissions. In response, the Dutch Climate Agreement (2019) established a target to reduce net CO₂ emissions from fen meadows by 1 Mt CO₂ per year by 2030. To achieve this objective, the Dutch National Research Programme on Greenhouse Gases in Peatlands (NOBV) implemented an intensive monitoring network that integrates chamber-based measurements with both on-site and airborne eddy covariance (EC) observations. A primary challenge in this context is attributing and upscaling CO₂ emissions across diverse peat types, soil conditions, groundwater regimes, and grassland management practices.

Direct measurement of peat oxidation at the ecosystem scale is not feasible; instead, it must be inferred from EC fluxes that encompass autotrophic respiration, heterotrophic decomposition, and management-induced vegetation turnover resulting from mowing and regrowth. Attribution remains challenging because conventional emission–response functions emphasise groundwater levels while neglecting soil physical properties and vegetation dynamics, which is a significant limitation in highly degraded, nutrient-rich peatlands.

Our modelling strategy consists of two components. First, we employ machine learning to analyse the data without imposing prior assumptions. Shapley-based attribution quantifies the contributions of groundwater depth, meteorological forcing, vegetation state, and mowing timing to fluxes, along with their interactions. These models identify nonlinear thresholds and regime-dependent behaviours that are challenging to specify a priori. We compare response structures across sites to assess sensitivities, rather than prescribing specific management scenarios.

Second, we develop a physics-based deep learning framework that integrates biophysically meaningful equations with adaptable learning components. Groundwater dynamics regulate oxygen availability and determine the proportion of organic matter exposed to air. Soil moisture profiles are used to characterise oxic and anoxic zones. Bulk density and porosity, which define degraded peat, constrain oxygen diffusion and moisture retention. By explicitly representing vegetation growth and mowing disturbances, we distinguish autotrophic respiration from peat oxidation.

Integrating soil physics and groundwater dynamics within a deep learning framework enhances both temporal robustness and cross-site transferability while maintaining model flexibility. This approach enables inference of peat oxidation from EC observations in a mechanistically consistent manner, thereby providing a robust foundation for evaluating mitigation strategies in intensively managed peatlands.