Abstract: Global lake methane emissions constitute a substantial flux from inland waters to the atmosphere. Recent model upscaling efforts have demonstrated the potential magnitude and heterogeneity of these emissions. Multiple sources of uncertainty can, however, become magnified in global extrapolations if not explicitly addressed. As upscaling efforts advance, it is critical to synthesize empirical measurements of methane fluxes from lakes, and assess how different factors in the measurements and models affect the predictability of global lake methane emissions. We conducted a model exercise that assessed how different uncertainty sources affected the total variability of global lake methane emission predictions and identified key areas for improvement in order to create more robust estimates of current and future emissions.
We collated global data of directly observed lake diffusive and ebullitive methane emissions, including 1166 and 611 separate methane diffusion and ebullition fluxes, respectively. We applied a Bayesian model framework to calibrate modified Arrhenius temperature models on half of the observed diffusion and ebullition observations, validating the model with the remaining observations. We then generated a posterior prediction using our calibrated model to the Global Lake, Climate, and Population database to predict out-of-sample methane diffusion and ebullition rates across 1.4+ million lakes at least 10 ha in size. Our predictions used one-at-a-time sensitivity analyses and Markov Chain Monte Carlo to represent sources of uncertainty, including model process, model parameter, observation, and driver data uncertainty, by making multiple draws from estimated distributions to build an ensemble of possible upscaling estimate outcomes.
Here we show that the posterior predictions of global lake methane emission estimates exhibit up to four orders of magnitude in total prediction uncertainty. The highlight of this study is the well-constrained and partitioned uncertainty, rather than the mean estimate which is similar to current global estimates. At first glance this uncertainty may seem uninterpretable, however, when partitioned we show that model process uncertainty (i.e., uncertainty in the model structure) was the largest source of total prediction uncertainty, a factor currently unaccounted for in existing global upscale estimates. The next highest contribution to uncertainty was from observation uncertainty followed by parameter uncertainty. In summary, this modeling exercise emphasizes that more robust global estimates of methane emissions from lakes can be generated by explicitly assessing how different uncertainty sources affect the total variability of global predictions.
