Using explainable machine learning to study restored peatland CH4 flux heterogeneity

Eddy Covariance (EC) method provides a valuable opportunity to monitor greenhouse gases, enabling informed decisions on climate change mitigation. Despite the abundance of EC data, explainable machine learning (ML) methods have not been effectively utilized to study the complex nature of methane (CH₄) fluxes, especially heterogeneity of emissions within ecosystems. This study explores the application of random forest ML model to analyse CH₄ flux spatiotemporal heterogeneity using flux data from the Ess-soo restored peatland in Estonia. This site, 30 years ago abandoned peat extraction area, was restored in 2021. To study CO₂ and CH₄ fluxes, open path EC analysers (LI-7500 and LI-7700, LICOR Biosciences) were installed in 2023. Additionally, CO₂ and CH₄ fluxes were measured biweekly using chamber method with the LI-7810 trace gas analyser (LICOR Biosciences) from 12 sampling spots in the EC footprint area. Other parameters such as water pH, electrical conductivity, dissolved oxygen concentration, temperature, oxidation reduction potenital, pH, and water level were conducted.

Chamber measurements revealed significant spatial CH₄ heterogeneity within EC flux footprint. The mean CH₄ flux from chamber measurement points during the summer months was 0.052 ± 0.013 µmol m^-2 s^-1 with a range of -0.001 to 0.555 µmol m^-2 s^-1. Looking into whole year EC dataset, main driver for CH₄ flux was water temperature. Day and nighttime fluxes responded differently to environmental changes, with air temperature and wind speed being significant drivers for day and night, respectively. The random forest model predicted CH₄ heterogeneity considerably better than general linear models performed (R² = 0.31 and 0.10, respectively). Besides identifying the main drivers, ML models can also combine EC and chamber measurements to detect hotspots and moments that are overlooked by EC alone. In that case, high spatial or temporal resolution  remote sensing data (e.g. LiDAR, Sentinel-1, Sentinel-2) was used. For instance, topographic wetness index calculated from LiDAR data in all points within EC flux footprint, was combined with water level—an important driver of both EC and chamber CH₄ fluxes. This information, together with chamber data was used to train ML models to estimate CH₄ fluxes spatially and temporally.

This work brings out the advantages in using ML and high spatial and temporal resolution remote sensing data to study CH₄ flux heterogeneity in wetlands. However, more testing is needed to see if these methods give similar results in other wetland sites.