Upscaling CO2 fluxes from agricultural drained lowland peatlands in England using remote sensing and machine learning

Abstract

Drained lowland peatlands in the UK are used as prime agricultural areas but are significant sources of CO₂ emissions. Monitoring and quantifying CO₂ dynamics in these ecosystems is critical to achieving the UK’s legal net-zero target by 2050. This study pioneers the upscaling of carbon fluxes (Gross Ecosystem Productivity (GEP), Total Ecosystem Respiration (TER), and Net Ecosystem Exchange of CO₂ (NEE)) in East Anglia’s agricultural peatlands (England) using remote sensing (RS) and machine learning (ML). A Random Forest model, trained with Landsat and Sentinel-2 imagery, meteorological data, and soil carbon information, predicts field-scale CO₂ fluxes with 77% overall accuracy. TER prediction was the strongest (R² = 0.84; RMSE = 1.18 gC/m²/d; NRMSE = 8%), followed by NEE (R² = 0.77; RMSE = 1.37 gC/m²/d; NRMSE = 8.13%), and GEP (R² = 0.76, RMSE = 1.97 gC/m²/d; NRMSE = 9.87%). The average predictive uncertainty for 14-day fluxes was $\pm 1.69$ gC/m²/d, which scaled with magnitude. The model was more accurate in grasslands compared to croplands. We validated the model with spatial cross-validation, finding it accurately predicts NEE seasonality at an unseen grassland site but deviates from observed mean values in winter and spring. We demonstrate the applicability of the model by upscaling annual and seasonal fluxes across the Fens, where the annual NEE in 2023 ranged from 1.04 to -2.52 kgC/m², depicting high spatial variability. This study establishes a baseline NEE scenario for the Fens and lays the groundwork for refining CO₂ flux modelling in drained peatlands, highlighting the potential of RS and ML for supporting the UK’s GHG reduction strategies in peatland ecosystems.