Near-term, iterative forecasting suggests high predictability of reservoir methane ebullition at weekly time scales

Abstract:

Near-term, iterative ecological forecasting with data assimilation is a useful approach for improving our understanding of biogeochemical processes. Most ecological forecasting workflows use automated, high-frequency sensor data to iteratively update predictions in near-real time as new observations become available. However, many ecosystem-level biogeochemical processes cannot yet be quantified reliably with automated sensors and instead depend on manually collected samples and laboratory analysis, representing a major gap in our understanding of the predictability of some ecosystem-level processes. One important biogeochemical process that near-term, iterative ecological forecasts could potentially improve estimates of – yet is challenging to both monitor and predict – is freshwater methane (CH₄) ebullition. We developed a forecasting and data assimilation workflow that generated real-time CH₄ ebullition rate forecasts for up to 16 days into the future. The forecast parameters and states were updated as manually collected ebullition data became available approximately weekly, and we quantified and partitioned the total forecast uncertainty among different sources (parameters, driver data, model process, and initial conditions). Our near-term, iterative forecast workflow successfully predicted an ensemble of future CH₄ ebullition rates that encompassed observations. Forecast skill increased and total forecast uncertainty decreased with each iterative forecasting cycle, and forecasts performed better than a null deterministic model in replicating observations. Importantly, our forecasts better predicted the total seasonal observed CH₄ ebullition than the deterministic model, which overestimated total emissions by 48%. The dominant contributor to total forecast uncertainty changed from parameters during a training period to model process and driver data during the forecasting period. Despite the high variability of CH₄ ebullition emissions in freshwater ecosystems, our iterative forecasting workflow demonstrates that ebullition can be successfully forecasted and has high predictability on approximately weekly timescales.