Defensible gap filling for eddy-covariance fluxes.
Screen EddyPro data, align CIMIS meteorology, compare Reichstein-style MDS with Random Forest and XGBoost using contiguous blocked validation, then create a provenance-rich LE/H product.
1. Load field data
EddyPro full-output TXT is the primary input.
2. Processing policy
Transparent defaults; measured data are never overwritten.
Quality, completeness and gap structure
Screened observations and predictor availability after timestamp regularization and optional CIMIS alignment.
Measured turbulent fluxes
Downsampled only for visualization; calculations use every record.
Project details
Predictor availability
Tower and external-reference features visible to ML models.
| Variable | Complete | N |
|---|
Validate models on realistic calendar outages
Unlike random CV or RF out-of-bag scores, records are hidden in contiguous windows and scored only against observations that were actually available before masking.
Validation design
Quick mode uses one window per duration. Increase windows for a stronger site-specific benchmark.
Adaptive fill policy
Choose the best validated method for each gap-duration class, or force one method.
Blocked-validation performance
| Flux | Gap | Model | RMSE | MAE | Bias | R² | Windows |
|---|---|---|---|---|---|---|---|
| Run validation to populate this table. | |||||||
Selected model by duration
Gap-filled fluxes with full provenance
Measured observations remain unchanged. Filled values are labeled by method and linked to the closest blocked-validation performance estimate.
Final LE and H
Method provenance
Daily ET from final LE
Temperature-dependent latent heat of vaporization.
Take the complete audit trail with the data.
Exports are generated in your browser. The compact product contains final fluxes, source labels, gap lengths and validation diagnostics. The full product retains the screened EddyPro/CIMIS columns plus all gap-fill fields.