MARS-Hyperspectral dataset

This repository contains the data for the work first publicly presented as a poster at the:

[Poster] Růžička, Mateo-García, Irakulis-Loitxate et al., Machine Learning Models for Multi-sensor Detection of Methane Leaks in Hyperspectral Data, June. 23-27, 2025, ESA Living Planet Symposium 2025.

Abstract

The Methane Alert and Response System (MARS), managed by the International Methane Emissions Observatory (IMEO) is a global satellite based initiative to detect and notify oil and gas methane emissions to governments and companies to swift mitigation action. Among the different satellites that MARS utilises, hyperspectral sensors such as EMIT, PRISMA and EnMAP offer an optimal trade-off between methane absorption sensitivity and spatial resolution that enables the attribution of methane plumes to individual facilities. Scanning large archives of satellite imagery to find methane plumes is a challenging task because matched filter retrievals still produce a significant amount of artifacts. Since April 2025 MARS is aided by AI models to detect potential plumes in hyperspectral images of EMIT, PRISMA and EnMAP. These detections are displayed within a QA/QC tool where analysts log in and verify or reject the detections. If the detection is recent and can be attributed to a facility in the ground a formal notification is issued to government and operators and MARS case-managers engage with the parties to trigger mitigation action.

Simply put, this repository contains the datasets used in training and evaluation of the deployed models.

Status

The paper and full repository are currently in development. More updates will follow soon!

Usage

Here we provide geotif files containing datasets built from three hyperspectral sensors (EMIT, EnMAP and PRISMA). These are split by the main folders, each sensor also contains a full-tile dataset (used for full-tile evaluation). Each of these then contains a folder per tile (either 256x256 or the full granule) corresponding to methane leak events and background locations. These are split into train/(sometimes)val/test splits using the csv files.

Each data sample has the following geo-located files (we recommend using for example QGIS to inspect these):

• plumemask.tif - binary mask (0 and 1) containing the main event label
• rgb.tif - RGB bands (extracted as single band from the hyperspectral data), also contains validity
• wmf.tif - wide matched filter following https://doi.org/10.5194/amt-17-1333-2024 (namely the "ΔXCH4(SWIR)" variant)
• mag1c.tif - mag1c matched filter following https://doi.org/10.1109/TGRS.2020.2976888 (note: not computed for all samples)
• mf.tif - basic matched filter product
• info.json - metadata about each event or background tile, notably for EMIT it contains the name of the original .nc file. Note that the 'isplume' information here refers to if the source was a plume, or other known locations (for example identified confounder areas, prior rejected plumes etc.) - as such it shouldn't be used as a flag whether the binary label contains any positive pixels (for that we recommend loading the plumemask.tif instead).
• (full tiles only) plumeinstances.tif - separated instances of multiple plumes in the single full tile (for easier visualisation)