CarbonFluxBench: A Global Benchmark for Upscaling of Carbon Fluxes Using Zero-Shot Learning

CarbonFluxBench comprises over 1.3 million daily observations from 573 eddy covariance flux tower sites globally (2000–2024). It provides stratified evaluation protocols that explicitly test generalization across unseen vegetation types and climate regimes, a harmonized set of remote sensing and meteorological features, and reproducible baselines ranging from tree-based methods to domain-generalization architectures.

Paper: CarbonFluxBench (KDD 2026) Code: github.com/alexxxroz/CarbonFluxBench

Dataset Summary

Property	Value
Daily observations	1,405,813
Flux tower sites	573
Date range	2000–2024
Source networks	FLUXNET2015, AmeriFlux, ICOS, JapanFlux
IGBP vegetation classes	16
Köppen climate classes	5 (main) / 30 (detailed)

Data Files

File	Description	Size
target_fluxes.parquet	Carbon flux targets + site metadata	~43 MB
MOD09GA.parquet	MODIS MOD09GA surface reflectance features	~474 MB
ERA5.parquet	ERA5-Land meteorological features	~5 GB
koppen_sites.json	Site → Köppen climate classification mapping	~11 KB
feature_sets.json	ERA5 feature set definitions (minimal/standard/full)	~7 KB
FLUXNET2015_Metadata.csv	FLUXNET2015 site metadata	~18 KB
AmeriFlux_Metadata.tsv	AmeriFlux site metadata	~210 KB
ICOS2025_Metadata.csv	ICOS site metadata	~3 KB

Prediction Targets

All targets are derived from eddy covariance measurements standardized under the ONEFlux methodology (units: gC m⁻² day⁻¹):

Target	Column	Description
GPP	GPP_NT_VUT_USTAR50	Gross Primary Production
RECO	RECO_NT_VUT_USTAR50	Ecosystem Respiration
NEE	NEE_VUT_USTAR50	Net Ecosystem Exchange (NEE = −GPP + RECO)

Each observation includes a continuous quality control flag: NEE_VUT_USTAR50_QC (0–1).

Features

MODIS MOD09GA (12 features)

Seven surface reflectance bands (sur_refl_b01–sur_refl_b07), sensor/solar geometry (SensorZenith, SensorAzimuth, SolarZenith, SolarAzimuth), and cloud fraction (clouds).

ERA5-Land (6 / 36 / 150 features)

Three configurable feature sets defined in feature_sets.json:

• Minimal (6): temperature, precipitation, radiation, evaporation, LAI (high & low vegetation)
• Standard (36): minimal + soil temperature/moisture (4 levels), wind, pressure, snow/albedo, radiation components, runoff
• Full (150): standard + lake variables, additional flux components, min/max daily variants

Site Metadata (5 features)

Latitude, longitude, IGBP vegetation type (16 classes), Köppen climate class (5 main / 30 detailed).

Train-Test Splits

CarbonFluxBench provides two complementary site-holdout splits for zero-shot evaluation (random state = 56):

• IGBP-stratified: partitioned by vegetation type. 80/20 for common classes (>10 sites), 50/50 for rare classes (≤10 sites).
• Köppen-stratified: partitioned by climate zone. Uniform 80/20 split across 5 main classes.

All splits are at the site level — train and test sites are mutually exclusive.

Usage

Download

pip install huggingface_hub
huggingface-cli download alexroz/CarbonFluxBench --repo-type dataset --local-dir data

With the CarbonFluxBench package

import carbonfluxbench

targets = ['GPP_NT_VUT_USTAR50', 'RECO_NT_VUT_USTAR50', 'NEE_VUT_USTAR50']
y = carbonfluxbench.load_targets(targets, include_qc=True)
y_train, y_test = carbonfluxbench.split_targets(y, split_type='Koppen')

modis = carbonfluxbench.load_modis()
era = carbonfluxbench.load_era('minimal')
train, val, test, x_scaler, y_scaler = carbonfluxbench.join_features(
    y_train, y_test, modis, era, scale=True
)

Direct loading with pandas

import pandas as pd

targets = pd.read_parquet("data/target_fluxes.parquet")
modis = pd.read_parquet("data/MOD09GA.parquet")
era5 = pd.read_parquet("data/ERA5.parquet")

Evaluation

All metrics are computed per-site, then reported as quantiles (25th, median, 75th percentile):

Metric	Description
R²	Coefficient of determination
RMSE	Root mean squared error (gC m⁻² day⁻¹)
nMAE	Mean absolute error normalized by site mean flux
RAE	Relative absolute error

Citation

License

MIT