AMERI-FAR25 Dataset for Carbon Flux Mapping

• AMERI-FAR25 is a dataset combining flux tower measurements with Landsat imagery to provide high-resolution carbon flux data across North American ecosystems.
• It comprises 7.7 million half-hour records from 209 AmeriFlux sites spanning 2013–2023, enabling detailed spatial and temporal analyses.
• Its use in the FAR deep-learning framework demonstrates an effective method for upscaling point measurements to pixel-level ecosystem carbon mapping.

The AMERI-FAR25 dataset is a publicly available resource constructed for high-resolution carbon flux prediction across diverse North American ecosystems. It pairs eddy-covariance tower flux measurements with co-located Landsat 8 and 9 imagery, forming the basis of the Footprint-Aware Regression (FAR) deep-learning framework for pixel-level ecosystem carbon flux estimation at 30 m spatial resolution (Searcy et al., 1 Dec 2025).

1. Dataset Composition and Scope

AMERI-FAR25 comprises data from 209 distinct AmeriFlux sites contributing a total of 439 site-years between 2013 and 2023. The geographical range includes the United States, Canada, Mexico, and Peru. The dataset contains 7,697,145 half-hourly records of net ecosystem carbon flux (FC), spanning a variety of ecosystem types as defined by IGBP codes:

• DBF (deciduous broadleaf forest)
• ENF (evergreen needleleaf forest)
• EBF (evergreen broadleaf forest)
• WET (wetlands)
• GRA (grasslands)
• CSH/OSH (shrublands)
• CRO (croplands)
• SAV (savannas)
• MF (mixed forests)
• CVM/BSV and additional minor classes

Each sample is spatially represented as a 128×128 pixel Landsat scene (≈3.84 km × 3.84 km), centered on the tower location.

Dimension	Value	Ecosystem Coverage
Sites	209	DBF, ENF, WET, GRA, CRO, etc.
Site-years	439	2013–2023 (varies by site)
Flux records	7,697,145	Half-hour, spatially co-located

2. Data Types and Variables

The dataset integrates flux, meteorological, and remote-sensing inputs:

Tower-Measured Variables:

• Net ecosystem exchange (FC) at half-hour intervals (Mg C ha⁻¹ half-hour⁻¹)

Meteorological/Environmental Drivers:

• For footprint modeling (X_footprint): wind direction (WD), wind speed (WS), friction velocity (USTAR), air temperature (TA), sensible heat flux (H), tower height.
• For flux prediction (X_drivers): shortwave incoming radiation (SW_IN), air temperature (TA), relative humidity (RH)
• PRISM normals (800 m grid): daily TA, RH; monthly solar transmission; SW_IN estimated via pysolar

Satellite Inputs:

• Landsat 8 & 9 bands resampled to 30 m: coastal aerosol, blue, green, red, NIR, SWIR1, SWIR2, cirrus, TIRS1, TIRS2—excluding the panchromatic band.
• Ancillary: sun/sensor azimuth and zenith
• Cloud-filtered scenes: 45,124 valid patches

3. Spatial and Temporal Resolution

AMERI-FAR25 delivers high spatial precision:

• Pixel size: 30 m × 30 m (thermal bands resampled from native 100 m)
• Patch size: 128 × 128 pixels (~4 km side length)
• Tower sampling: every 30 minutes
• Landsat revisit: nominally every 16 days; "most recent available" scene per flux record
• Aggregation for model evaluation: monthly and annual sums of flux

Footprint modeling leverages soft attention-based masks $F_P \in \mathbb{R}^{128 \times 128}$, normalized such that $\sum_{l=1}^{128} \sum_{w=1}^{128} F_{P,lw}(t)=1$. Pixel-level predictions for $FC_{pixel}(l,w,t)$ are aggregated using this mask:

$$\hat y_t = \sum_{l=1}^{128} \sum_{w=1}^{128} \bigl[FC_{pixel}(l,w,t)\times F_{P,lw}(t)\bigr]$$

4. Preprocessing and Quality Control

Tower Data Cleaning:

• Inclusion: AmeriFlux BASE, CC-BY-4.0 license, required meteorological and flux variables
• SW_IN derived from PPFD_IN via linear scaling where necessary
• Metadata harmonized (highest sensor, tower height required)
• Outlier removal: FC outside 0.5–99.5% percentiles; negative SW_IN; night drawdown exclusion (FC<0 with SW_IN=0)

Satellite Imagery Processing:

• Download via landsatxplore with custom reliability fixes
• QA/QC using PIXEL_QA flags (clear land/water: codes 21824, 21888, 21952)
• Missing/cloudy pixels filled via temporal back-fill from latest valid observation

Co-Registration:

• Patches centered on tower coordinates; resampled bands; orientation angles concatenated as four supplementary channels

5. Data Splitting and Modeling Protocols

Spatial Splitting:

• Sites grouped by IGBP ecosystem class
• For classes with ≥10 sites: 40% withheld (20% validation, 20% test)
• Classes with <10 sites: used exclusively for training

Temporal Splitting:

• val_future/test_future: final year from multi-year sites, targeted at temporal-drift studies
• val: 20% random holdout of remaining half-hour records

Split	Tower Records	Landsat Patches
Training (train)	~5.6 million	~35,000
Validation/Test (site)	~2.1 million	~10,000
val/test (future year)	~0.5 million	~5,000

A plausible implication is that these splits facilitate ecosystem- and temporally robust generalization studies for upscaling models.

6. Metadata and Annotations

AMERI-FAR25 includes extensive metadata relevant for both scientific reproducibility and ecological interpretation:

• Land cover and ecosystem codes (IGBP type per site)
• Tower metadata: height, coordinates, operation years, documented disturbance events (e.g., clear-cut, fire scar)
• Sun/sensor azimuth and zenith per scene (enabling bidirectional reflectance modeling)

Data are made available on Zenodo (DOI pending) with FAR code, model weights, and full site list published at github.com/jsearcy1/FAR and a corresponding archive.

7. Context, Intended Use, and Implications

AMERI-FAR25 underpins the Footprint-Aware Regression (FAR) framework, which achieves an $R^2 = 0.78$ for monthly net ecosystem exchange prediction on holdout test sites. The dataset allows pixel-level upscaling of ground-validated fluxes, addressing the mismatch between tower and satellite spatial scales. Researchers may apply AMERI-FAR25 directly for landscape-scale, high-resolution carbon flux mapping in heterogeneous environments, facilitating cross-ecosystem analyses and method development for natural climate solutions (Searcy et al., 1 Dec 2025).