Modified

27 October 2026

Level-4.1: Gap-filling

The gaps of each USTAR scenario (CUT_16, CUT_50, CUT_84) are filled with XGBoost, the method of this dataset, and with MDS, the standard method of FLUXNET, for comparison only. Use the XGBoost series. The gap-filled series end in _gfXG and _gfMDS; their flags FLAG_..._ISFILLED are 0 where the flux was measured.

Method, evidence and checks:

Input

The gap-filling starts from the records that passed Levels 2 to 3.3 with a quality flag of 0 or 1, by day and by night (NEE_L3.1_L3.3_<scenario>_QCF, see QCF). The highest-quality records, flag 0, are in NEE_L3.1_L3.3_<scenario>_QCF0; use them for analyses that need measured values only.

The drivers are three complete meteo series at 47 m: air temperature (TA_T1_47_1_HOMOGENIZED_gfXG), vapour pressure deficit (VPD_T1_47_1_HOMOGENIZED) and incoming shortwave radiation (SW_IN_T1_47_1_gfXG). Temperature and VPD are the homogenised versions, which correct for the probe changes in 2016 and June 2026.

XGBoost

Each year has its own model for each scenario, trained on that year and its two neighbouring years, with features engineered from the drivers and selected with SHAP values. All scenarios use the same features. The number of trees is chosen once by cross-validation on held-out blocks of 14 days (up to 1000 tested) and used for every year: 110 trees for NEE, 130 for LE and 140 for H. The gaps are filled by a model refitted on all complete records. The 2016-2024 processing used a random forest.

MDS

MDS (mode='diive') fills every gap, also long ones, and uses only measured driver values as similar conditions. In long gaps it can rest on very few similar records, so a few unusual values near the edge of a gap set the fill for days. The notebook splits the difference between the sums of MDS and XGBoost by day and night, month and gap length.

Long gaps without measurements

The model fills these gaps without measurements nearby:

  • 1 to 22 January 2016, before and during the installation of the LI-7200. With the start-up period removed by hand (1 March to 25 May), 2016 lacks about 108 days of measurements, including the leaf-out, so its annual sum rests largely on the model;
  • 18 May to 7 June 2018, failed flow module;
  • 10 January to 20 March 2019, both LI-7200 units defective, then the LI-7500 as a stand-in, whose fluxes are in progress for FP2026.2 and not part of FP2026.1;
  • 18 December 2020 to 4 January 2021;
  • 13 December 2021 to 7 February 2022, raw data that cannot be converted.

LE has the same gaps as FC, and the periods removed by hand add to them. H exists in several periods without FC (the 2016 start-up, 18 May to 7 June 2018, 6 to 10 January 2019, 18 December 2020 to 4 January 2021); its long gaps are 10 January to 20 March 2019 and 13 December 2021 to 7 February 2022. LE and H have no USTAR scenarios, no storage term and no uncertainty columns.

Uncertainty

The uncertainty of the CUT_50 series of each method follows Pastorello et al. (2020), in µmol m-2 s-1 at 1 sigma:

Column Content
_RANDUNC Random uncertainty: spread of measured fluxes under similar conditions
_USTARUNC USTAR uncertainty: half the difference between the CUT_84 and CUT_16 series
_JOINTUNC Both combined: √(RANDUNC² + USTARUNC²)

For annual sums the random part adds in quadrature, the USTAR part as half the difference between the CUT_84 and CUT_16 sums. The difference between XGBoost and MDS is not part of the joint uncertainty; the notebook reports it beside it.

Known limitations

  • MDS in spring 2016. The data from 1 March to 25 May 2016 are removed (start-up of the LI-7200, see Level 3.2). MDS fills this gap from time windows with a median width of 70 days and gives far more uptake than XGBoost: April and May 2016 sum to -235 gC m-2 with MDS against -113 gC m-2 with XGBoost (CUT_50). In the other years, April and May range from -30 to -195 gC m-2 with either method. Use the XGBoost series.
  • H from 10 January to 20 March 2019 is model output only: there is no Level-1 H in this period (see Level 1).
  • H in 2024 is about a third lower than in the other years. This has not been checked yet.
Back to top