Integrating machine learning with a process-based model for estimating global wetland methane emissions

Chris C R Smith, Shuo Chen, Sparkle L Malone, Gavin McNicol, Qing Zhu, Licheng Liu, Youmi Oh

Estimates of methane emissions from natural wetlands are uncertain and depend on the modeling approach. Process-based models incorporate knowledge of the underlying biogeochemistry, but prediction accuracy is sensitive to parameterization and empirical equations. Machine learning models have potential to improve estimates, however they struggle to generalize to new prediction sites. We explore combined process-based machine learning strategies. Prediction accuracy improved when two new inputs were incorporated: the estimate from a process-based surrogate model and satellite imagery from each measurement site. We estimated global-scale emissions using conditional model selection: a machine learning model was applied to grid cells within the range of environmental conditions used for training, and a process-based model was used for more extreme regions. The hybrid approach resulted in a global emissions estimate of 180.9 teragrams of methane per year, which was larger than either of the individual models.