F
FluxGapFillUNIVERSAL IMPORT · QC · MDS · RF · XGBOOST · PHASE 2
Files stay local
Loading Python
Phase 2 · quality control, gap filling and blocked validation

Bring your processed flux data.
FluxGapFill maps the rest.

Start from computed fluxes or half-hourly/hourly environmental data—not raw high-frequency sonic/IRGA signals. Phase 2 adds configurable QC, two MDS variants, enhanced RF/XGBoost candidates, realistic blocked validation, empirical uncertainty calibration and adaptive gap filling.

EddyProCampbell TOA5AmeriFlux / FLUXNETCSV / TXT / TSVExcelManual mapping fallback100% browser processing

1. Project data

Primary processed flux/environmental file plus optional supplemental weather data.

No alternate-format files to test?Use the bundled synthetic examples:
Campbell TOA5FLUXNET-styleGeneric CSVExcelSupplemental weather

2. Project policy

Time handling is explicit because logger conventions differ.

Privacy behavior. Selected files are copied only into the temporary Pyodide filesystem/memory for this browser session. Refreshing or closing the page removes the working copies; original files on your computer are unchanged.
Variable mapping

Verify before processing

FluxGapFill proposes mappings and units but does not silently assume unknown fields. Correct any suggestion here before loading the project.

Choose files and run inspection first.

Primary file

Not inspected

—
RoleSource columnUnitDetection
Inspect a primary file first.

Supplemental file

Optional

—
RoleSource columnUnitDetection
No supplemental file selected.

Input preview

First records only. Preview does not alter the file.

Inspect a file to preview it.
Data overview

Quality, completeness and gap structure

Standardized midpoint grid after mapping, unit normalization, QC screening and optional supplemental alignment.

Records—
Time step—
LE missing—
H missing—
LE longest gap—
H longest gap—
Period—
External source—

Measured fluxes

Downsampled only for visualization; calculations use every record.

Project details

Project—
Primary source—
Period—
Records—
Resolution—
Supplemental source—

Predictor availability

Variables visible to current/future analysis modules.

VariableCompleteN
Phase 2 · quality control

Screen transparently before model training

Phase-1 input/QC screening remains intact. This additional layer can reject implausible ranges and isolated robust spikes without altering the original standardized copy.

QC policy

Conservative defaults; every rejection receives an explicit reason flag.

Reversible within the session. Re-running QC always starts from the standardized Phase-1 import, not from the previously screened values.

QC summary

Upstream input/QC rejection plus Phase-2 screening remain distinguishable.

Range rejects—
Spike rejects—
Variables—

QC audit

No rejected value is silently removed; the full provenance export retains pre-QC values and flags.

VariableBeforeRange rejectSpike rejectAfterMissing after
Load a project and apply Phase-2 QC.
Capability-aware workspace

Use what is ready. Add data where needed. Skip the rest.

Each analysis module checks required variables independently, so missing footprint or carbon information never blocks gap filling or ET work.

Optional site information

Needed by selected advanced modules; leave blank if not applicable.

How optional modules behave

Ready means the required data are already in the project. Needs data identifies exactly what is missing. Use “Add data” to return to the import workspace, or “Skip” to hide that module for this project. Skipping never deletes data.
High-frequency raw EC processingOut of scope
Processed flux / environmental dataSupported
Server-side storageNone
Process a dataset to evaluate module readiness.
Phase 2 · model lab

Compare established MDS with enhanced machine learning

Validation hides contiguous calendar outages and scores only observations that existed before masking. This is retrospective reconstruction validation—not forecasting.

Blocked validation design

Use multiple windows for stronger site-specific evidence. Long runs can take several minutes in the browser.

Adaptive final fill

The best validated method is selected by target and nearest tested gap duration. You can also force one candidate.

Uncertainty is calibrated separately. MDS donor SD is retained as donor dispersion. A 95% reconstruction interval is only generated when enough blocked-validation residuals exist.

Overall blocked-validation performance

RMSE, MAE, bias, R² and prediction coverage across the hidden calendar windows.

FluxGapModelNRMSEMAEBiasR²Windows
Run validation to populate this table.

Day / night diagnostics

Day/night basis appears after validation.

FluxGapModelSegmentNRMSEBias
Run validation first.

Selected model by duration

Site-specific adaptive rules from the validation results.

No site-specific validation has been run yet.

Seasonal diagnostics

DJF, MAM, JJA and SON are reported when a held-out window contains observations from that season.

FluxGapModelSeasonNRMSEMAEBiasR²
Run validation first.
Results

Gap-filled fluxes with provenance

Measured observations remain unchanged; reconstructed values retain method labels, gap class, MDS donor diagnostics and empirical blocked-validation uncertainty when available.

Final LE and H

Intervals: run blocked validation before final filling to calibrate empirical uncertainty.

Method provenance

Run gap filling to populate provenance.

Daily ET from final LE

Temperature-dependent latent heat of vaporization.

Phase 2 · reproducible export

Take the complete audit trail with the data.

All exports are generated locally in the browser. No result file is uploaded to FluxGapFill.

Gap-filled products

QC & validation diagnostics

Preparing scientific runtime

Loading Python…