Upscaling CO2 fluxes from the UK's agriculturally drained peatlands using Remote Sensing and Machine Learning

Drained peatlands are responsible for 5.6% of global anthropogenic CO₂ emissions, yet the conventional algorithms for quantifying CO₂ fluxes are not well-calibrated and validated within these ecosystems. In the UK, drained peatlands serve as key agricultural areas but account for approximately 24% of the country’s peatland emissions. Reducing emissions from agriculturally drained peatlands is a vital component of the UK’s net zero strategy, and monitoring CO₂ dynamics in these ecosystems is essential for meeting net zero targets by 2050. To support these efforts, we evaluate the potential of remote sensing data integrated with machine learning methods to upscale carbon fluxes (GEP, TER, and NEE) measured by eddy covariance flux towers in agriculturally-drained peatlands of the Fenland, UK, for the first time. We used moderate-resolution data from Landsat and Sentinel 2 in combination with meteorological parameters and soil carbon data to train a Random Forest model capable of predicting CO₂ fluxes at the field scale. The model showed an overall accuracy of 77\%, with an R² of 0.81 and RMSE of 2.23 kgCO₂/m²/yr for predicting partitioned fluxes. NEE, calculated as the difference between modeled GEP and TER achieved an R² of 0.78 and RMSE of 1.61 kgCO₂/m²/yr. The model showed the highest predictive accuracy in managed grasslands and showed weaker performance in the arable site on deep peat and specific crop types (e.g., sugar beet and leek). On an unseen eddy covariance site, the model effectively captured the seasonal pattern of NEE but showed deviations from observed seasonal averages in winters (+0.75 kgCO₂/m²/yr) and spring (+1.42 kgCO₂/m²/yr). We demonstrate the applicability of the model by upscaling field-level annual and seasonal fluxes across the Fenland, where the average NEE in 2023 showed high spatial variability (ranging from 3.79 to -9.2 kgCO₂/m²). This work enables the creation of a baseline NEE scenario for any field of interest within lowland peatlands of the UK, which can be monitored over time to evaluate the efficacy of restoration efforts, such as partial or complete rewetting of grasslands, as well as the impact of changes in management practices. Overall, this assessment establishes a foundation for advancing CO₂ flux modeling in drained peatlands and demonstrates the potential of remote sensing and machine learning approaches to support greenhouse gas (GHG) mitigation efforts in the UK’s peatland ecosystems.