Metamodel simulation of carbon fluxes across an eroding and pristine blanket bog in Scotland

Abstract

Accurately quantifying carbon dynamics in peatlands is critical to assess their role in regulating global climate. Within hotspots of peatland degradation, such as in Europe and South-east Asia, skilful assessment of the spatial and temporal impacts of climate change and different land management options is required to meet emissions reductions targets and improve regional management planning. To address this challenge, a random forest-based metamodel was evaluated to assess its utility in simulating various greenhouse gas (CO2) emission components, including Net Ecosystem Exchange (NEE), Gross Primary Productivity (GPP), and Ecosystem Respiration (ER) across two Scottish peatlands. The metamodel mimicked the complex Wetland-DNDC model at a higher level of abstraction with increased efficiency and lower computational time. While Wetland-DNDC also simulates NEE, GPP and ER, it typically involves a considerable number of parameters related to soil properties, climate data, vegetation characteristics, biogeochemical processes, hydrology, nutrient cycling, and microbial activity. Many of these parameters (more than 100) are challenging to measure in the field, and literature values are often adopted, which may not necessarily reflect local site conditions. In essence, this multidimensional parameter space introduces high uncertainties in modelling carbon fluxes. In contrast, random forest-based metamodel preserved the key relationships between NEE and input variables (air and soil temperature, water table, precipitation, vegetation, and soil properties) as described in the Wetland-DNDC model with lower parameter requirements (less than 20) and increased accuracy. Similar unique relationships were established for GPP and ER. The random forest-based metamodel represented the Wetland-DNDC model  within the spectrum of input values and parameters across which it was simulated. The simulation was conducted in two locations across Scotland with contrasting contemporary carbon dynamics: a near natural blanket bog in Cross Lochs, Forsinard, currently functioning as a resilient net carbon dioxide sink (UK-CLS; Lat. = 58.37, Long. = -3.96; altitude = 207 m) and an eroding oceanic blanket bog located in the Cairngorms, currently net emitting carbon dioxide (UK-BAM; Lat. = 56.92, Long. = -3.15, altitude = 642 m). The simulation was validated against eddy covariance flux measurements under varying climate conditions. In contrast to Wetland-DNDC (R2 = 0.43), the metamodel provided a much-improved fit to the 1:1 line for NEE (R2 = 0.83). Model accuracy was slightly lower for the former (RMSE = 0.72) compared to its metamodel version (RMSE = 0.699). Similar trends were observed for GPP and ER simulations. At a monthly resolution, Wetland-DNDC-derived NEE, GPP, and ER consistently deviated by more than 20% from the eddy covariance-derived estimates, whereas its metamodel version showed deviations of less than 10%. Currently, work is in progress to incorporate management and drought simulation within a metamodel framework, as well as to upscale carbon fluxes from tower to landscape resolution. The simulation of carbon fluxes using the metamodel-based approach holds the promise of enhancing emission reporting to Tier 3 standards and offers a hopeful avenue for modelling carbon dynamics in peatlands.