Multi-Decadal Dynamics of Wetland Methane Emissions Revealed by Knowledge-Guided Machine Learning.

Area of Science:

• Environmental Science
• Climate Science
• Biogeochemistry

Background:

• Methane flux (FCH₄) measurements from wetlands are limited compared to carbon dioxide, hindering long-term analysis and climate feedback assessments.
• Extrapolating short-term wetland FCH₄ data is difficult for both process-based and machine learning (ML) models.
• Understanding wetland methane dynamics is critical for climate change research.

Purpose of the Study:

• To develop a knowledge-guided ML framework for reconstructing long-term wetland methane budgets and trends.
• To integrate eddy covariance (EC) FCH₄ observations, warming experiments, and biogeochemical knowledge.
• To provide robust, long-term datasets for validating ecosystem models and advancing wetland biogeochemistry understanding.

Main Methods:

• Developed a novel knowledge-guided machine learning (ML) framework.
• Integrated eddy covariance (EC) methane flux observations from 11 AmeriFlux sites.
• Incorporated field warming experiment data and biogeochemical knowledge.

Main Results:

• Reconstructed multi-decadal wetland FCH₄ trends from 2000-2024.
• Observed significant variability in FCH₄ trends, with increases up to 14% per decade.
• Found that increasing FCH₄ trends weaken from high to low latitudes, indicating northern wetland vulnerability.

Conclusions:

• The study provides novel, robust reconstructions of long-term wetland methane fluxes.
• Findings highlight the vulnerability of northern wetlands and latitudinal variations in methane trends.
• The generated datasets serve as critical benchmarks for ecosystem models and improve understanding of wetland biogeochemistry.