Machine Learning-based Methane Flux Projections at selected sites in the Northern Hemisphere.

Abstract:

Climate change is expected to impact methane (CH4) fluxes from ecosystems, which can create a positive feedback loop. Eddy covariance (EC) measurements at ecosystems scale aim to detect CH4 flux changes and trends. However, EC measurements included in the FLUXNET-CH4 database have few years of data, which has so far constrained trend detection. Here, we used data from 9 FLUXNET-CH4 sites to train a machine learning model. Sites with at least three years of data and with a coefficient of determination higher than 65 were selected, including: Fens (4), Marsh (1), Bog (1), Wet Tundra (1), Rice Paddy (1), and a subtropical freshwater marsh (1). The model enabled us to reconstruct CH4 fluxes since 1980 (based on atmospheric data from MERRA2 reanalysis), and project future CH4 fluxes by using simulations from four selected CMIP5 Global Climate Models. We found that, under a high-emission scenario (RCP8.5), increases in CH4 fluxes expected for end of century range from about 80% at the Rice Paddy (likely due to the higher air temperature and longwave radiation), to about 10% at the sub-tropical site (attributable to warmer and dryer summer conditions). The 30-day rolling mean of air temperature was the most important predictor.