Gap-filling carbon dioxide, water, energy, and methane fluxes in challenging ecosystems: Comparing between methods, drivers, and gap-lengths

Summary

This study extends gap-filling methodology for eddy covariance flux measurements by comparing random forest regression and traditional approaches (notably marginal distribution sampling) across challenging ecosystems including managed pastures, converted peatlands, and drylands. Whilst MDS remains preferable for short gaps in carbon dioxide data, random forest regression demonstrated superior performance for longer gaps and for all non-CO₂ fluxes (methane, sensible heat, latent energy), with particular value when measured environmental drivers are unavailable and when filling gaps exceeding three months.

UK applicability

Findings are directly applicable to UK managed grassland and pasture systems where eddy covariance monitoring is increasingly deployed. The recommendation that random forest regression outperforms traditional methods for longer gaps and methane flux estimation will be valuable for UK peatland and upland pasture flux research and monitoring programmes.

Key measures

Gap-filling accuracy for carbon dioxide, water, energy, and methane fluxes; cumulative flux reconstruction; performance across gap lengths (short <12 days, medium 12-30 days, long >30 days, very long >3 months); ecosystem-flux response preservation

Outcomes reported

The study evaluated random forest regression (RFR) and marginal distribution sampling (MDS) approaches for filling missing data in eddy covariance measurements of carbon dioxide, water, energy, and methane fluxes across challenging ecosystems including managed pastures, converted peatlands, and drylands. Performance was assessed across different gap lengths and ecosystem types to determine which methods most reliably reconstructed cumulative fluxes.

Topic tags