MethaneSET: Unified Multi-Sensor Datasets for Satellite-Based Methane Plume Detection

MethaneSET is a collection of AI-ready datasets for satellite-based methane plume detection spanning three sensors (Sentinel-2, Landsat-8/9, and EMIT) with unified metadata. It provides labeled scenes with expert-verified plume segmentation masks, unlabeled plume-free scenes for self-supervised learning, a synthetic plume bank from WRF-LES simulations for data augmentation, and a MODTRAN-based transmittance look-up table for physics-informed plume injection.

Paper: MethaneSET: Unified Multi-Sensor Datasets for Satellite-Based Methane Plume Detection (Scientific Data, under review)

Authors: Cesar Aybar, Julio Contreras, David Montero, Miguel D. Mahecha, Luis Gómez-Chova

Dataset Components

Component	Sensor	Description	Samples	Size
methaneset-s2-finetune	Sentinel-2	Labeled scenes with plume masks	~5,500	~45 GB
methaneset-s2-pretraining	Sentinel-2	Confirmed plume-free scenes	~55,000	~450 GB
methaneset-l89-finetune	Landsat-8/9	Labeled scenes with plume masks	~5,500	~17 GB
methaneset-l89-pretraining	Landsat-8/9	Confirmed plume-free scenes	~55,000	~150 GB
methaneset-emit	EMIT	Radiance cubes + MF products + masks	~1,500	~1.2 TB
methaneset-bank	Synthetic	WRF-LES column-integrated CH₄ maps	1,064,448	~15 GB
integrated_transmittances.json	—	MODTRAN6 transmittance LUT	—	<1 MB

Total: ~2.0 TB

Annotation Sources

Plume labels come from two independent operational monitoring systems that provide consensus verification through cross-referencing.

IMEO MARS (International Methane Emissions Observatory, Methane Alert and Response System) processes satellite imagery to detect methane plumes globally. For multispectral sensors it uses multi-band multi-pass (MBMP) enhancement maps followed by manual analyst review. For EMIT it applies matched-filter retrievals with structured quality control. The system has identified over 5,500 verified plumes across 1,315 sites in 22 countries.

Carbon Mapper operates an independent detection pipeline using adaptive matched-filter retrievals on EMIT and Tanager-1 data. Their four-level processing system includes automated retrieval, analyst screening, quality flagging, and final verification. The portal contains over 9,000 published plume observations.

For the EMIT component, MethaneSET includes only scenes where both IMEO and Carbon Mapper independently confirmed a plume, providing high-confidence consensus labels (~70% overlap between the two systems across ~4,000 EMIT plumes from each source).

Multispectral Components (Sentinel-2 / Landsat-8/9)

Detection Paradigm

Multispectral methane detection uses temporal image pairs. Methane absorbs strongly in the SWIR2 band (2190 nm, B12 for S2, B7 for Landsat) and weakly in SWIR1 (1610 nm). By comparing a target acquisition against a plume-free reference from a nearby date, the MBMP ratio isolates the absorption signal from surface reflectance changes.

Assets per Sample

Asset	Description
target	Target acquisition (13 bands for S2, 11 bands for L89)
reference	Cloud-free reference image (same bands, ±90 days)
ch4	CH₄ enhancement map (ppb) from MARS-S2L
plume	Binary segmentation mask (expert-verified)

All images are 200×200 pixels at 10 m resolution (S2) or 10 m resampled (L89), stored as Cloud-Optimized GeoTIFFs with Zstandard compression.

Labeled vs Unlabeled Split

The finetune subsets contain scenes with confirmed plumes and their segmentation masks. The pretraining subsets contain scenes from the same monitored sites where no plume was detected on that date, confirmed through the MARS-S2L operational pipeline. These plume-free scenes support self-supervised pretraining and serve as clean backgrounds for synthetic plume injection.

Metadata Fields

All metadata is queryable via SQL through the TACO Parquet catalog without loading any imagery.

satellite:platform      — S2A, S2B, S2C, LC08, LC09
satellite:sza           — Solar Zenith Angle [degrees]
satellite:vza           — View Zenith Angle [degrees]
satellite:saa           — Solar Azimuth Angle [degrees]
datetime                — ISO 8601 acquisition timestamp
geometry                — WKB footprint
methane:fluxrate        — Estimated emission rate [kg/h] (labeled only)
methane:source_type     — Emission source category
site:country            — ISO country code
site:id                 — MARS site identifier

Hyperspectral Component (EMIT)

Detection Paradigm

EMIT (Earth Surface Mineral Dust Source Investigation) is a NASA imaging spectrometer on the ISS with 285 contiguous bands (381–2493 nm) at 60 m resolution. Unlike multispectral sensors, hyperspectral instruments can resolve individual CH₄ absorption features from a single acquisition using matched-filter techniques, eliminating the need for temporal reference imagery.

Two detection approaches exist for hyperspectral data. Retrieval-based methods apply a matched filter to calibrated radiance to produce a methane enhancement map, then use the enhancement map as input to a segmentation model. End-to-end methods operate directly on calibrated radiance cubes, learning to detect plumes without an explicit retrieval step. MethaneSET supports both approaches.

Assets per Sample

Asset	Description
radiance	Calibrated L1B radiance cube (285 bands, sensor coordinates)
matched_filter	Standard matched-filter enhancement product
mag1c	Mag1c matched-filter retrieval
wmf	Wide-window matched-filter retrieval
plume_sensor	Binary segmentation mask (sensor coordinates)
plume_geo	Binary segmentation mask (geographic coordinates)
glt	Geometric Lookup Table for georeferencing

Radiance cubes and matched-filter products are stored in native sensor coordinates to preserve exact band-to-band registration. Plume masks are provided in both sensor coordinates (aligned with radiance for training) and geographic coordinates (for visualization and GIS). The GLT arrays are included to enable on-demand georeferencing of any product.

Selection Criteria

The EMIT component includes approximately 1,500 scenes where plumes were independently confirmed by both IMEO MARS and Carbon Mapper. This dual-source requirement ensures high label confidence at the cost of excluding scenes detected by only one system.

Synthetic Plume Bank

The plume bank contains 1,064,448 pre-computed column-integrated CH₄ enhancement maps generated from Large Eddy Simulations using WRF-LES (Gorroño et al., 2025). These simulations model turbulent plume dispersion at 20 m resolution, capturing realistic atmospheric dynamics that Gaussian approximations cannot reproduce.

Parameter Space

Parameter	Values	Count
Wind speed	2.0, 2.5, 3.0, 3.5, 4.0, 4.5, 5.0, 6.0, 7.0, 8.0, 10.0 m/s	11
Source configuration	Point and area sources	14
Turbulent snapshot	Independent realizations	3
Wind direction	0° to 330°, step 30°	12
Solar Zenith Angle	0° to 70°, step 10°	8
Solar Azimuth Angle	0° to 345°, step 15°	24

All simulations use a reference emission rate of 3,000 kg/h. Enhancement maps scale linearly to simulate different source strengths during injection.

Plume Injection Pipeline

The companion methaneset package provides a physics-based injection pipeline. Given a clean background scene, the pipeline queries the bank for the plume best matching the observation geometry (SZA, SAA, wind speed and direction), scales the enhancement to a target emission rate, and applies transmittance corrections via the MODTRAN6 look-up table. The LUT maps (Air Mass Factor, CH₄ concentration) to band-integrated transmittance per sensor variant, capturing the non-linear absorption that a simple Beer-Lambert formulation cannot represent. The complete pipeline executes in approximately 5 ms per sample, enabling on-the-fly augmentation during training.

Quick Start

import tacoreader

tacoreader.use("pandas")

# Load any component — instant, no bulk download
ds = tacoreader.load(
    "https://huggingface.co/datasets/tacofoundation/methaneset"
    "/resolve/main/methaneset-l89-finetune/"
)

# Spatiotemporal filtering on the Parquet index
subset = ds.filter_datetime("2022/2023").filter_bbox(-104, 31, -102, 33)

# Read a single sample — imagery fetched on demand via HTTP range requests
sample = subset.read(0)
target = sample.read("target")      # local path to GeoTIFF
plume  = sample.read("plume")       # local path to mask

print(f"Samples in subset: {len(subset.data)}")

SQL Queries Across Sensors

# All plumes above 1000 kg/h in Iraq — works across any sensor
ds.sql("SELECT * FROM level0 WHERE methane:fluxrate > 1000 AND site:country = 'IQ'")

Synthetic Plume Injection

from methaneset import MethaneBank, TransmittanceLUT, inject_plume

bank = MethaneBank("methaneset-bank/")
lut  = TransmittanceLUT("integrated_transmittances.json")

result = inject_plume(
    background=swir_bands,    # (2, H, W) array [B12, B11]
    lut=lut, bank=bank,
    sza=35.0, saa=None,
    sensor="LC08",
    q_target=5000.0,          # kg/h
)
# result["image"]  — modified SWIR bands
# result["mask"]   — ground-truth enhancement map

TACO Format

All components follow the TACO specification (Aybar et al., 2025). Each dataset consists of a Parquet catalog containing all metadata (coordinates, timestamps, sensor parameters, emission rates, site information) alongside Cloud-Optimized GeoTIFFs for the imagery. The Parquet catalog can be queried with SQL via DuckDB without downloading or opening any image file, enabling rapid filtering across hundreds of thousands of samples. See tacoreader documentation for details.

License

MethaneSET is released under CC-BY-4.0. The synthetic plume bank and transmittance LUT carry no additional restrictions. EMIT L1B radiance data is sourced from NASA LP DAAC (public domain). Sentinel-2 and Landsat data are sourced from Copernicus and USGS respectively (open access). Plume annotations are derived from IMEO MARS and Carbon Mapper operational products.

Citation

@article{aybar2026methaneset,
  title     = {MethaneSET: Unified Multi-Sensor Datasets for Satellite-Based Methane Plume Detection},
  author    = {Aybar, Cesar and Contreras, Julio and Montero, David and Mahecha, Miguel D. and G{\'o}mez-Chova, Luis},
  journal   = {Scientific Data},
  year      = {2026},
  note      = {Under review}
}

Acknowledgments

We thank the United Nations Environment Programme International Methane Emissions Observatory (UNEP-IMEO) for providing access to the MARS-S2L and MARS-Hyperspectral operational databases, and the Carbon Mapper team for their publicly available plume observations. The synthetic plume bank is based on WRF-LES simulations by Gorroño et al. (2025). EMIT data are courtesy of the NASA Jet Propulsion Laboratory and distributed by the LP DAAC.