Contribution of wavelets to decrease gap filling in turbulent surface fluxes measurements

There is an urgent need to provide data for greenhouse gases (GHG) monitoring. Flux towers currently represent the most direct approach to provide continuous datasets of surface exchange of these gases. Typically, the flux is derived from the EC method over 30 minutes averaging periods. However, technical and meteorological conditions create gaps with missing or non-reliable data which can represent from 20 % to 60 % of the total period. To compute annual GHG balances, one needs to fill these gaps. Studies on benchmark forest datasets show that gap-filling methods, such as the marginal distribution sampling (MDS), have a good performance on hourly and daily data as well as on annual budgets. This overall setup (EC and MDS gap-filling) is the standard for consolidated datasets over FLUXNET and most flux networks (e.g. Warm Winter 2020 available at www.icos-cp.eu/).

Direct observations are the base of the gap-filling itself and should be praised over gap-filling techniques. More so, the flux community nowadays faces new challenges by moving towards less ideal sites (e.g. European PAUL-ICOS-Cities project to monitor city GHG fluxes), and facing increasing extremes conditions (e.g. Drought-2018, available at www.icos-cp.eu/). In these less conventional frameworks, gap-filling techniques need to be evaluated. Among flux processing techniques, wavelets (WV) can reliably measure non-stationary periods and thus retain more data than standard EC. It does that by resolving flux calculation at sampling rate. For flux measurements it has been most notably used for airborne measurement as it allows computing a flux over short enough periods to attribute the measured flux to a limiter land area.

Here we analyse the CO2 flux over the three consecutive years (2019, 2020 and 2021) in two ICOS sites: FR-Gri, a crop site and FR-Fon a mixed deciduous forest site, both near Paris, France. Results show that from 52 606 data points in FR-Fon, around 50% needed to be gap filled, this includes 11% of missing data and 22% of periods with developed turbulence but non-stationary. Preliminary results for 2019 in FR-Gri show similar ranges. By not requiring stationarity, WV method avoid close to half of the gap filling compared to EC. Comparing 30-minutes-averaged fluxes derived from EC and WV shows good correlation (R²=0.99 for observed data and 0.94 for gap filled data), low root mean square error (RMSE=1.12  for observed data and 1.95 for gap filling). The extra data also decreased continuous gaps, which is expected to improve the performance of gap filling methods. More so, during summer 2019 heatwaves stroke Europe and in particularly French ecosystems. On the flux data, daytime observations during elevated temperatures in 2019 show a WV derived CO₂ fluxes closer to zero, suggesting the expected response of stomatal closure during these events. The gap-filled EC data, however, showed relatively unchanged photosynthesis. This study shows the usefulness of using WV computed CO₂ fluxes, a result expected to remain valid for longer time series and for other ecosystems and meteorological conditions.