Global net biome CO₂ exchange predicted comparably well using parameter-environment relationships and plant...

Area of Science:

• Ecology
• Climate Science
• Computational Modeling

Background:

• Accurate net biome CO2 exchange (NBE) estimation is crucial for understanding terrestrial ecosystems' role in climate change.
• Previous NBE prediction improvements focused on model complexity, but parametric uncertainty significantly impacts model skill.
• Traditional plant functional type (PFT)-based parameterization assigns uniform parameters within land cover regions.

Purpose of the Study:

• To investigate how different parameterization assumptions affect NBE prediction errors globally.
• To compare a novel machine learning-based 'environmental filtering' (EF) approach against the traditional PFT-based method.
• To isolate the impact of parametric uncertainty on NBE predictions using a model-data fusion framework.

Main Methods:

• Simulated PFT-based and EF-based parameterization within the CARDAMOM model-data fusion framework at a global scale.
• EF approach predicts pixel parameters based on relationships with climate, soil, and canopy properties.
• Benchmarked PFT and EF predictions against CARDAMOM's Bayesian optimization approach for 'true' parameters.

Main Results:

• The EF approach matched or outperformed the PFT approach at 55% of pixels in terms of mean absolute error of NBE predictions.
• EF-based NBE estimates showed susceptibility to error compensation between component fluxes.
• Predicted parameters in the EF approach sometimes aligned poorly with assumed 'true' values.

Conclusions:

• The EF approach is comparable to conventional methods for NBE prediction and warrants further investigation.
• Understanding parametric uncertainty is key to improving terrestrial biosphere model performance.
• This research informs the development of PFT-free and trait-based model parameterization strategies.