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

Methane emissions from natural wetlands are a major contributor to the changing global climate. However, estimates of such emissions are uncertain and depend on the modeling approach used. Process-based modeling of wetland emissions incorporates scientific knowledge of the underlying biogeochemical process, but prediction accuracy is insufficient. Machine learning models have the potential to improve emissions estimates; however, they often struggle to generalize to new prediction sites. We explore combined process-based machine learning strategies for estimating wetland emissions globally. One such strategy, cross-domain model stacking — using the output from a process-based surrogate model as an additional input to a second neural network trained on real data — improved reducing prediction error by 2.1%, while other transfer learning methods did not improve performance. In addition, we find that learning features from MODIS satellite images local to each measurement site served as a valuable input to the neural network, improving performance by 3.7%. Finally, we estimate global-scale emissions using a new, hybrid approach: conditional model selection dependent on local environment. The hybrid approach resulted in a global emissions estimate of 180.9 teragrams of methane per year, which is larger than either of the individual models, although the global sum was strongly affected by the inundation map. Last, integrating process-based information with machine learning improved alignment with established trend and seasonal behaviors while maintaining or enhancing agreement with independent machine learning-based global estimates.