pith. sign in

arxiv: 2509.00626 · v7 · pith:UZSDEIACnew · submitted 2025-08-30 · 💻 cs.CV · cs.AI

Towards Methane Detection Onboard Satellites

Maggie Chen , Hala Lamdouar , Luca Marini , Laura Mart\'inez-Ferrer , Chris Bridges , Giacomo Acciarini This is my paper

Pith reviewed 2026-05-18 19:02 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords methane detectionmachine learningsatellite imageryhyperspectral imagesunorthorectified dataplume detectionEMIT sensoronboard processing
0
0 comments X

The pith

Machine learning detects methane plumes from unorthorectified satellite images at performance comparable to corrected data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that machine learning models can identify methane plumes directly from raw, unorthorectified hyperspectral images captured by the EMIT sensor. Conventional approaches first correct geometric distortions through orthorectification and then apply matched filters to strengthen plume signals, but the authors demonstrate these steps are not required for effective ML detection. Models trained on the unorthorectified dataset reach similar accuracy to those trained on orthorectified versions. When orthorectified data is used, the ML models exceed the performance of the standard mag1c matched filter baseline. This approach opens the possibility of running detection algorithms onboard satellites to enable faster responses while cutting the amount of data that must be downlinked.

Core claim

The central claim is that ML models trained on unorthorectified EMIT hyperspectral images achieve detection performance comparable to models trained on orthorectified images, while models trained on orthorectified images outperform the mag1c matched filter baseline.

What carries the argument

UnorthoDOS, the dataset of unorthorectified hyperspectral images from EMIT used to train ML plume detectors without orthorectification or matched filtering.

If this is right

  • Detection can run directly on satellite hardware without prior geometric correction.
  • Onboard ML reduces the volume of raw data that must be transmitted to the ground.
  • ML trained on orthorectified EMIT data surpasses the mag1c matched filter for plume detection.
  • The released datasets enable training of additional models for rapid methane monitoring.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same bypass of preprocessing could apply to detection of other trace gases in hyperspectral satellite streams.
  • Lower data volumes from onboard detection would reduce ground station load and allow more frequent observations.
  • Testing the models on imagery from different hyperspectral sensors would show whether the approach generalizes beyond EMIT.

Load-bearing premise

Unorthorectified images retain enough methane plume signal for machine learning to learn despite geometric distortions.

What would settle it

A head-to-head test in which ML models trained on unorthorectified images show substantially lower detection accuracy than models trained on orthorectified images would disprove the central claim.

Figures

Figures reproduced from arXiv: 2509.00626 by Chris Bridges, Giacomo Acciarini, Hala Lamdouar, Laura Mart\'inez-Ferrer, Luca Marini, Maggie Chen.

Figure 1
Figure 1. Figure 1: Schematic representation of the unorthorectified data generation process. We refer to the resulting unorthorectified dataset as the Unorthorectified Dataset for On Board Satellite methane detection (UnorthoDOS) [10]. Orthorectified benchmark For comparison, an orthorectified dataset was generated using the orthorectified counterparts of the same L1B images used in the unorthorectified dataset, while retain… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the detection pipeline. EMIT images are preprocessed into orthorectified and unorthorectified datasets using spatial jittering as data augmentation. The ML models provide an onboard early detection system by performing classification for the TIP satellite and segmentation for the CUE satellite. The datasets are split to reserve a holdout test set of tiles from 5% of L1B images, with the rest di… view at source ↗
Figure 3
Figure 3. Figure 3: Visualisation of semantic segmentation results on 2 example tiles from the orthorecti￾fied 3a and the unorthorectified 3b datasets each. From Left to Right in each sub-figure: L1B tiles (only RGB bands shown for visualisation) overlaid with ground truth methane plume annotations; predicted semantic segmentation plume masks from UNet; segmentation predictions from mag1c [4] establishing that plume detection… view at source ↗
read the original abstract

Methane is a potent greenhouse gas and a major driver of climate change, making its timely detection critical for effective mitigation. Machine learning (ML) deployed onboard satellites can enable rapid detection while reducing downlink costs, supporting faster response systems. Conventional methane detection methods often rely on image processing techniques, such as orthorectification to correct geometric distortions and matched filters to enhance plume signals. We introduce a novel approach that bypasses these preprocessing steps by using \textit{unorthorectified} data (UnorthoDOS). We find that ML models trained on this dataset achieve performance comparable to those trained on orthorectified data. Moreover, we also train models on an orthorectified dataset, showing that they can outperform the matched filter baseline (mag1c). We release model checkpoints and two ML-ready datasets comprising orthorectified and unorthorectified hyperspectral images from the Earth Surface Mineral Dust Source Investigation (EMIT) sensor at https://huggingface.co/datasets/SpaceML/UnorthoDOS , along with code at https://github.com/spaceml-org/plume-hunter.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces the UnorthoDOS dataset of unorthorectified hyperspectral images from the EMIT sensor and claims that ML models trained on this data achieve performance comparable to models trained on orthorectified versions. It further states that models trained on orthorectified data outperform the mag1c matched-filter baseline. The work releases two ML-ready datasets, model checkpoints, and code with the goal of enabling onboard methane plume detection without orthorectification or matched filtering.

Significance. If the empirical results hold with rigorous validation, the finding that convolutional or transformer models can localize plumes from raw, geometrically distorted EMIT cubes would be significant for onboard processing, as it could eliminate computationally expensive preprocessing steps and reduce downlink volume. The open release of both orthorectified and unorthorectified ML-ready datasets plus reproducible code is a clear strength that supports community follow-up and deployment.

major comments (2)
  1. [Abstract] Abstract: the central claims of 'comparable performance' for unorthorectified training and outperformance of mag1c are stated without any quantitative metrics (precision, recall, F1, IoU, or AUC), without test-set size or scene selection criteria, and without a description of the validation procedure or cross-validation scheme. This absence makes the soundness of the primary empirical result impossible to evaluate from the available text.
  2. [Dataset Description] Dataset construction: it is not stated how plume labels were transferred or re-projected from orthorectified annotations onto the unorthorectified cubes. If labels were copied without accounting for shear, stretch, or band-to-band misalignment, the reported parity between unorthorectified and orthorectified training could be an artifact of label noise rather than evidence of learned robustness to geometric distortion.
minor comments (2)
  1. [Experimental Setup] The manuscript would benefit from an explicit statement of the exact train/validation/test split ratios and the number of unique plumes or scenes in each partition.
  2. [Figures] Figure captions should include the precise definition of the performance metric plotted (e.g., whether it is per-pixel or per-plume) and the number of test samples used.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback. We address each major comment point-by-point below. Where the comments identify gaps in clarity or completeness, we have revised the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claims of 'comparable performance' for unorthorectified training and outperformance of mag1c are stated without any quantitative metrics (precision, recall, F1, IoU, or AUC), without test-set size or scene selection criteria, and without a description of the validation procedure or cross-validation scheme. This absence makes the soundness of the primary empirical result impossible to evaluate from the available text.

    Authors: We agree that the abstract would be strengthened by including quantitative support for the claims. In the revised version we have updated the abstract to report specific metrics (F1-score and IoU) for the unorthorectified versus orthorectified models and for the comparison against mag1c, along with the size of the held-out test set and a brief statement that evaluation was performed on scenes not seen during training. The full experimental protocol, including scene selection and validation details, remains in the Methods and Experiments sections. revision: yes

  2. Referee: [Dataset Description] Dataset construction: it is not stated how plume labels were transferred or re-projected from orthorectified annotations onto the unorthorectified cubes. If labels were copied without accounting for shear, stretch, or band-to-band misalignment, the reported parity between unorthorectified and orthorectified training could be an artifact of label noise rather than evidence of learned robustness to geometric distortion.

    Authors: We thank the referee for highlighting this important detail. The unorthorectified labels were generated by mapping the orthorectified plume annotations onto the corresponding unorthorectified cubes using the EMIT sensor's per-pixel geolocation metadata (latitude/longitude and viewing geometry). We have added a new paragraph in the Dataset Construction section that describes this projection step, notes the handling of residual geometric distortions, and reports that label fidelity was verified by visual inspection on a random subset of scenes. This addition makes the label-transfer procedure explicit and addresses the concern that performance parity might be an artifact of label noise. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical ML training on released datasets

full rationale

The paper reports standard supervised training of ML models (CNNs/transformers) on two released EMIT hyperspectral datasets (orthorectified and unorthorectified) and compares their plume-detection performance to the mag1c matched-filter baseline. No mathematical derivation, fitted parameters renamed as predictions, or self-citation chain is invoked to justify the central claims. Results are obtained by direct experimentation on external sensor data with released code and checkpoints, satisfying the self-contained benchmark criterion. No load-bearing step reduces to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The work rests on standard supervised learning assumptions (i.i.d. train/test splits, sufficient signal in raw pixels) and the quality of the public EMIT sensor data; no new physical axioms or invented entities are introduced.

axioms (1)
  • domain assumption Unorthorectified hyperspectral pixels contain sufficient spatial and spectral information for plume detection without explicit geometric correction.
    Invoked when claiming that models trained on UnorthoDOS achieve comparable performance.

pith-pipeline@v0.9.0 · 5736 in / 1189 out tokens · 43360 ms · 2026-05-18T19:02:23.173578+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages

  1. [1]

    Myhre, D

    G. Myhre, D. Shindell, F.-M. Bréon, W. Collins, J. Fuglestvedt, J. Huang, D. Koch, J.-F. Lamarque, D. Lee, B. Mendoza, T. Nakajima, A. Robock, G Stephens, T. Takemura, and H. Zhang. Anthropogenic and Natural Radiative Forcing. InClimate Change 2013 - The Physical Science Basis, pages 659–740. Cambridge University Press, 2014

    work page 2013

  2. [2]

    Detection algorithms in hyperspectral imaging systems: An overview of practical algorithms

    Dimitris Manolakis, Eric Truslow, Michael Pieper, Thomas Cooley, and Michael Brueggeman. Detection algorithms in hyperspectral imaging systems: An overview of practical algorithms. IEEE Signal Processing Magazine, 31(1):24–33, 2013

    work page 2013

  3. [3]

    Real-time remote detection and measurement for airborne imaging spectroscopy: a case study with methane.Atmospheric Measurement Techniques, 8(10):4383–4397, 2015

    DR Thompson, I Leifer, H Bovensmann, M Eastwood, M Fladeland, C Frankenberg, K Ger- ilowski, RO Green, S Kratwurst, T Krings, et al. Real-time remote detection and measurement for airborne imaging spectroscopy: a case study with methane.Atmospheric Measurement Techniques, 8(10):4383–4397, 2015

    work page 2015

  4. [4]

    Fast and accurate retrieval of methane concentration from imaging spectrometer data using sparsity prior.IEEE Transactions on Geoscience and Remote Sensing, 58(9):6480–6492, 2020

    Markus D Foote, Philip E Dennison, Andrew K Thorpe, David R Thompson, Siraput Jon- garamrungruang, Christian Frankenberg, and Sarang C Joshi. Fast and accurate retrieval of methane concentration from imaging spectrometer data using sparsity prior.IEEE Transactions on Geoscience and Remote Sensing, 58(9):6480–6492, 2020

    work page 2020

  5. [5]

    Semantic segmentation of methane plumes with hyperspectral machine learning models.Scientific Reports, 13(1):19999, November 2023

    Vít R˚ užiˇcka, Gonzalo Mateo-Garcia, Luis Gómez-Chova, Anna Vaughan, Luis Guanter, and Andrew Markham. Semantic segmentation of methane plumes with hyperspectral machine learning models.Scientific Reports, 13(1):19999, November 2023

    work page 2023

  6. [6]

    Hyperspectralvits: General hyperspectral models for on-board remote sensing.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2025

    Vít R˚ užiˇcka and Andrew Markham. Hyperspectralvits: General hyperspectral models for on-board remote sensing.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2025. 5

    work page 2025

  7. [7]

    Gabriele Meoni, Roberto Del Prete, Federico Serva, Alix De Beusscher, Olivier Colin, and Nicolas Longépé. Unlocking the use of raw multispectral earth observation imagery for onboard artificial intelligence.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17:12521–12537, 2024

    work page 2024

  8. [8]

    EMIT L1B At-Sensor Calibrated Radiance and Geolocation Data 60 m V001, 2022

    Robert Green. EMIT L1B At-Sensor Calibrated Radiance and Geolocation Data 60 m V001, 2022

    work page 2022

  9. [9]

    EMIT L2B Estimated Methane Plume Complexes 60 m V001, 2023

    Robert Green, Andrew Thorpe, Philip Brodrick, Dana Chadwick, Amanda Lopez, Clayton Elder, Claire Villanueva-Weeks, Jay Fahlen, Red Willow Coleman, Daniel Jensen, Holly Bender, Quentin Vinckier, Chuchu Xiang, Winston Olson-Duvall, Sarah Lundeen, and David Thompson. EMIT L2B Estimated Methane Plume Complexes 60 m V001, 2023

    work page 2023

  10. [10]

    UnorthoDOS - Unorthorectified Dataset for On board Satellite methane detection, 2025

    Maggie Chen, Hala Lambdouar, Luca Marini, Laura Martínez-Ferrer, Chris Bridges, and Giacomo Acciarini. UnorthoDOS - Unorthorectified Dataset for On board Satellite methane detection, 2025

    work page 2025

  11. [11]

    Greenhouse gas (ghg) emissions monitoring: Point source vs

    CEOS - Committee o nEarth Observation Satellites. Greenhouse gas (ghg) emissions monitoring: Point source vs. wide area mappers, Aug 2024. Accessed: 2025-08-18

    work page 2024

  12. [12]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InInternational Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015

    work page 2015

  13. [13]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. InInterna- tional Conference on Learning Representations (ICLR), 2015. 6

    work page 2015