2024 Abstracts

Integrating Process-Based and Machine Learning Approaches for Estimating the Global Methane Soil Sink

Authors

Youmi Oh1,2* ([email protected]), Licheng Liu3, Qing Zhu4, Gavin McNicol5, Sparkle Malone6, Zhenong Jin3

Institutions

1NOAA Global Monitoring Laboratory, Boulder, CO; 2Cooperative Institute for Research in Environmental Sciences, University of Colorado, Boulder, CO; 3University of Minnesota–Twin Cities, Minneapolis/St. Paul, MN; 4Lawrence Berkeley National Laboratory, Berkeley, CA; 5University of Illinois–Chicago, IL; 6Yale University, New Haven, CT

Abstract

Methane (CH4) voxidation by microbes is the second largest sink in the global CH4 budget, but the magnitude and variability in soil CH4 oxidation are uncertain. Recent studies identified an overlooked CH4 soil sink in diverse terrestrial ecosystems, attributed to high-affinity methanotrophs growing on atmospheric CH4 in dry mineral soils. The current estimation of the global methane soil sink is ~30 Tg yr-1 but with considerable uncertainty (7 to>100 Tg yr-1) from previous studies. Accurately quantifying the soil sink is vital to reduce the bias in current and future global CH4 budgets.

Process-based modeling and machine learning approaches have been widely used to quantify methane fluxes on regional and global scales, but both approaches show their limitations. Specifically, results show large uncertainties in process-based estimation due to parameter optimization and governing microbial processes. Results further show limitations of the machine learning approach due to out-of-sample prediction failure and low interpretability with key responses that govern soil CH4 oxidation processes.

In this context, the emerging field of Knowledge-Guided Machine Learning (KGML) offers a promising hybrid modeling method, integrating the strengths of process-based models, machine learning techniques, and multisource datasets. In this presentation, researchers introduce novel KGML framework, designed to incorporate biogeochemical knowledge into machine learning algorithms effectively. The framework integrates direct measurements from the FLUXNET-CH4 and chamber datasets, along with indirect measurements of global soil temperature and moisture, to train and validate the model. Through this innovative approach, researchers aim to enhance understanding of soil methane oxidation processes, reduce uncertainties in methane budgets, and foster more accurate projections of global CH4 dynamics.