Stable gap-filling for longer eddy covariance data gaps: A globally validated machine-learning approach for carbon dioxide, water, and energy fluxes

Continuous time-series of CO2, water, and energy fluxes are useful for evaluating the impacts of climate-change and management on ecosystems. The eddy covariance (EC) technique can provide continuous, direct measurements of ecosystem fluxes, but to achieve this gaps in data must be filled. Research-standard methods of gapfilling fluxes have tended to focus on CO2 fluxes in temperate forests and relatively short gaps of less than two weeks. A gap-filling method applicable to other fluxes and capable of filling longer gaps is needed. To address this challenge, we propose a novel gap-filling approach, Random Forest Robust (RFR). RFR can accommodate a wide range of data gap sizes, multiple flux types (i.e. CO2, water and energy fluxes). We configured RFR using either three (RFR3) or ten (RFR10) driving variables. RFR was tested globally on fluxes of CO2, latent heat (LE), and sensible heat (H) from 94 suitable FLUXNET2015 sites by using artificial gaps (from 1 to 30 days in length) and benchmarked against the standard marginal distribution sampling (MDS) method. In general, RFR improved on MDS's R2 by 15% (RFR3) and by 30% (RFR10) and reduced uncertainty by 70%. RFR's improvements in R2 for H and LE were more than twice the improvement observed for CO2 fluxes. Unlike MDS, RFR performed well for longer gaps; for example, the R2 of RFR methods in filling 30-day gaps dropped less than 4% relative to 1-day gaps, while the R2 of MDS dropped by 21%. Our results indicate that the RFR method can provide improved gap-filling of CO2, H and LE flux timeseries. Such improved continuous flux measurements, with low bias, can enhance our understanding of the impacts of climate-change and management on ecosystems globally.