Methane Detection On Board Satellites from Unorthorectified Imagery Maggie Chen∗1 , Hala Lamdouar∗1 , Luca Marini∗2 , Laura Martínez-Ferrer3 , Chris Bridges4 , Giacomo Acciarini5 1
University of Oxford, Oxford, United Kingdom Delft University of Technology, Delft, The Netherlands 3 Universitat de València, València, Spain 4 University of Surrey, Guildford, United Kingdom 5 European Space Agency, Advanced Concepts Team, Noordwijk, The Netherlands
arXiv:2609.04906v1 [cs.CV] 4 Sep 2026
2
As a potent greenhouse gas, methane is a major driver of climate change. Its effective mitigation relies on timely detection. Conventional detection methods rely on orthorectification to correct geometric distortions and matched filters to enhance plume signals, which are steps designed for ground processing and poorly suited to onboard execution. We introduce UnorthoDOS, a dataset and approach for training machine learning models directly on unorthorectified hyperspectral imagery, bypassing both orthorectification and matched-filter products. Our UNet models trained on unorthorectified data approach the performance of models trained on orthorectified data (IoU 16.91% vs. 18.47% on all plumes), while both substantially outperform the mag1c matched-filter baseline (IoU 4.76%). We further demonstrate the feasibility of onboard deployment: FP16 compression halves model size with under 0.3% output deviation. The trained ML models and two ML-ready datasets – orthorectified and unorthorectified hyperspectral imagery from the EMIT sensor – are publicly available at https://huggingface. co/datasets/SpaceML/UnorthoDOS, with code at https: //github.com/spaceml-org/plume-hunter.
1
Introduction
Methane has a global warming potential approximately 84 times greater than that of carbon dioxide over a 20year period [1]. A significant proportion of point-source methane emissions originates from “super-emitters” in the oil and gas sector, which are detectable from space [2]. As interest grows in deploying on board artificial intelligence (AI) for Earth observation (EO), real-time processing on satellites is becoming technically and operationally feasible [3–7]. When coupled with hyperspectral imaging, whose fine spectral resolution is well suited to identifying methane, on board AI provides a unique opportunity for rapid methane detection, reducing latency and downlink requirements [3, 8, 9]. This combination offers a pathway to faster response and more effective mitigation of ∗ These authors contributed equally and are listed in alphabetical order.
3rd Conference on AI in and for Space (SPAICE 2026) © 2026 M. Chen et al., licenced via CC BY 4.0.
high-impact methane releases. Methane plume detection in hyperspectral satellite imagery typically involves two key processing steps: (1) orthorectification, which corrects geometric distortions caused by sensor viewing angle, terrain variations, and Earth curvature, and (2) the generation of methane enhancement products, often using matched filters [10–12] to enhance weak plume signals by comparing against predefined spectral signatures. However, orthorectification is intended for ground processing and not designed on board execution [3, 13], and traditional matched filters are iterative algorithms with high computational cost and are susceptible to high false-positive rates [14]. Previous studies such as [15] improved segmentation accuracy by combining matched filter outputs with Red Green Blue (RGB) images using Deep Learning models. More recently, end-to-end approaches used lightweight Vision Transformers [16] to advance performance, and efficient, low-power algorithms have been proposed in [14]. Summarised in Table 1, these methods remain reliant on orthorectified data. This study introduces UnorthoDOS, an unorthorectified hyperspectral dataset and approach that better reflects on board acquisitions, and shows that methane plume detection trained directly on unorthorectified imagery achieves performance similar to that of orthorectified pipelines. It extends [17] with onboard model compression results, increasing the feasibility of deployment on resource-constrained hardware.
2
Methodology
2.1 Hyperspectral & Methane Data Source Datasets are constructed from hyperspectral observations acquired by the EMIT imaging spectrometer aboard the International Space Station (ISS), obtained via the NASA Earthdata portal1 . The L1B at-sensor radiance data product [18] provides 285 hyperspectral bands at 60 m spatial resolution, and the L2B methane plume complex product 1 https://search.earthdata.nasa.gov/
Dataset
Instrument
Bands
Spectral range (nm)
Unorthorectified
STARCOP [15]
AVIRIS-NG
125
RGB, 1573–1699, 2004–2480
✗
EMIT EMIT EMIT
86 86 86
RGB, 1573–1699, 2004–2478 RGB, 1573–1699, 2004–2478 RGB, 1573–1699, 2004–2478
✗ ✗ ✓
OxHyperSyntheticCH4 [16] OxHyperRealCH4 [16] UnorthoDOS (Ours)
Table 1: Comparison of hyperspectral datasets for methane plume detection, with emphasis on the presence or absence of orthorectification preprocessing.
[19], also at 60 m resolution, supplies orthorectified groundtruth plume masks. L2B annotations are available only in orthorectified form, which motivates the unorthorectification procedure described in Section 2.2. The full corpus spans 1,574 annotated methane plumes drawn from EMIT scenes acquired between August 10, 2022 and October 26, 2024. Following the band-selection procedure of [16], 86 of the 285 bands are retained, spanning 1573-1699 nm and 2004-2478 nm. This subset covers the primary methane absorption features while including a small number of nonabsorption bands and three RGB bands to let the model learn background spectral variability.
2.2
UnorthoDOS Dataset Generation
Iunortho −→ Iortho (x, y) 7−→ τ (x, y) = (xo , yo )
(1)
Unorthorectification is the inverse mapping τ −1 , approximated by reconstructing Iunortho from Iortho using a pixel coordinate grid in the source (unorthorectified) plane. The grid is then transformed forward through τ into the orthorectified plane, and used to sample pixels from the orthorectified methane plume annotations back into the unorthorectified plane. Because τ is not bijective, Iunortho and Iortho may differ in size. Consequently, the inverse mapping τ −1 is not necessarily surjective, leading to some pixels in Iunortho without corresponding values in Iortho . These missing values in Iunortho are filled using nearest-neighbor interpolation Φ. In summary, the procedure to generate an approximate off-nadir image can be expressed as: Iˆunortho = Φ ◦ τ −1 (Iortho ).
2.3
Orthorectified Benchmark Dataset
To isolate the effect of orthorectification on model performance, the orthorectified benchmark dataset is built from the same underlying EMIT scenes and 1,574 plumes as UnorthoDOS, using the orthorectified L1B radiance images paired with the original orthorectified L2B plume annotations and applying the tiling procedure described in Section 2.2.
2.4
The method to generate unorthorectified data is illustrated schematically in Figure 1. Orthorectification is the mapping τ that transforms pixel coordinates (x, y) from an angled (off-nadir) image Iunortho to coordinates (xo , yo ) in the orthogonal (nadir) image Iortho , compensating for sensor viewing geometry and terrain relief. This mapping is extracted directly from the geometric lookup table included in the EMIT L1B product. τ:
training. The resulting unorthorectified ML-ready dataset is referred to as UnorthoDOS (Unorthorectified Dataset for Onboard Satellite methane detection) [20].
Experimental Setup
To support a rapid detect-and-response pipeline, we adopt a simplified tip and cue paradigm [21] (Figure 2). A tip satellite performs lightweight binary classification to flag candidate methane plumes, and a cue satellite performs semantic segmentation to precisely localize them. Data splits are performed at the scene (L1B image) level, prior to tiling, to avoid tile leakage. Each of the orthorectified and unorthorectified datasets is split into a training set (80% of scenes), a validation set (15% of scenes), and a holdout test set (5% of scenes), with identical scene-level splits used across the two geometric variants. Given the limited number of annotated methane plumes, training and validation sets are augmented with random spatial translations. Input images values are normalised per spectral band by subtracting the training-set mean and dividing by the training-set standard deviation. A U-Net [22], adapted to accept 86-channel hyperspectral input, is used to predict binary per-pixel methane plume masks2 . A tile is classified as plume-positive if its predicted mask contains at least one positive pixel. Class imbalance between plume and background is addressed with the Dice loss [23]: LDice = 1 − D = 1 −
2T P . 2T P + F P + F N
(3)
(2)
Two U-Net models are trained separately on the orthorectified and unorthorectified datasets with Adam [24], learn-
Both hyperspectral images and their corresponding plume annotations are tiled into 128 × 128 pixel tiles for model
2 Full architecture and augmentation configuration are provided in the released code.
3rd Conference on AI in and for Space (SPAICE 2026) © 2026 M. Chen et al., licenced via CC BY 4.0.
Figure 1: Overview of the UnorthoDOS pipeline for synthesizing unorthorectified hyperspectral data from orthorectified EMIT data.
PREPROCESSING
TASKS
EARLY DETECTION SYSTEM
computed on an NVIDIA RTX 4070 Ti SUPER GPU (16 GB).
CLASSIFICATION Plume?
ORTHORECTIFIED DATA
Yes No
TIP CUE
Tip satellite
Plume detected!
Cue satellite
Prediction
UNORTHORECTIFIED DATA
Figure 2: Overview of the detection pipeline. EMIT images are preprocessed into orthorectified and unorthorectified datasets for ML model training. Trained ML models simulate a tip and cue setting to perform plume classification and segmentation.
ing rate 10−4 , batch size 32, for 100 epochs, selecting the best model with lowest Dice validation loss. Training is performed on a single 40 GB NVIDIA A100 GPU.
Performance Evaluation
The trained PyTorch models (prior to compression) are evaluated on the holdout test sets described in Section 2.4. We report precision, recall, and F1-score for both classification and segmentation, classification accuracy, and segmentation IoU, all computed at a 0.5 sigmoid threshold. To assess performance as a function of plume strength, we additionally report all metrics on a subset restricted to tiles whose maximum annotated methane concentration is ≥ 900 ppm m.
2.6
Results
SEGMENTATION Ground truth
2.5
3
Model Compression
Trained U-Net checkpoints are exported from PyTorch to the Open Neural Network Exchange (ONNX) format [25]. First, a ONNX graph optimization compresses models to 32-bit floating-point model (FP32) that is numerically equivalent to the original PyTorch checkpoint. This model is then further compressed to 16-bit floating-point precision (FP16), and separately to 8-bit integer precision (INT8) using weight-only quantisation. For each precision level, we report file size and the maximum per-pixel relative difference between compressed and original model outputs, 3rd Conference on AI in and for Space (SPAICE 2026) © 2026 M. Chen et al., licenced via CC BY 4.0.
Tables 2 and 3 report image classification and semantic segmentation performance for the U-Net models and the mag1c [12] baseline. For semantic segmentation, the orthorectified U-Net improves upon mag1c by 13.71 IoU points on all plumes (18.47% vs. 4.76%), and by 25.96 points on strong plumes (30.80% vs. 4.84%); the unorthorectified U-Net shows a comparable margin (16.91% vs. 4.76% and 26.17% vs. 4.84%, respectively). Qualitatively (Figure 3), U-Net predictions are visually coherent and closely follow plume boundaries, whereas mag1c produces fragmented outputs and a higher false-positive rate from misidentified terrestrial features, visible in the second unorthorectified example. Both segmentation and classification improve substantially when restricted to strong plumes, for both data regimes (Tables 2 and 3). Segmentation IoU rises from 18.47% to 30.80% (orthorectified) and from 16.91% to 26.17% (unorthorectified), while classification recall rises from 56.21% to 83.33% and from 48.89% to 71.61%, respectively. This pattern indicates that model performance is limited primarily by weak, low-concentration plumes rather than by a general failure to localize methane signal, a limitation we return to in Section 4. U-Net models trained on the unorthorectified and orthorectified data achieve comparable performances: IoU on all plumes is 16.91% vs. 18.47%, and on strong plumes 26.17% vs. 30.80%. Classification performance shows a wider gap: accuracy is 5.3 points lower on all plumes (71.48% vs. 76.80%) and 6.8 points lower on strong plumes (85.65% vs. 92.41%) for the unorthorectified case (Table 2). Despite this gap, both settings substantially outperform the mag1c baseline, and the unorthorectified model requires no orthorectification step at inference time. This constitutes, to our knowledge, the first demonstration that methane plume detection is feasible directly on unorthorectified satellite imagery, without the geometric correction step conventionally treated as a prerequisite. Model compression substantially reduces file size (Ta-
Setting
Model
Threshold (ppm m)
Precision
Recall
F1-Score
Accuracy (%)
Orthorectified
mag1c [12]
N/A ≥ 900
52.55 39.60
94.12 95.24
67.45 55.94
54.58 46.84
Orthorectified
U-Net
N/A ≥ 900
95.56 94.60
56.21 83.33
70.78 88.61
76.80 92.41
Unorthorectified
U-Net
N/A ≥ 900
89.19 87.88
48.89 71.61
63.16 78.91
71.48 85.65
Table 2: Image classification performance. Precision, recall, F1-score, and accuracy are reported for mag1c [12] and U-Net, trained/evaluated on orthorectified and unorthorectified data. Best result per column within each block is highlighted in bold.
Setting
Model
Threshold (ppm m)
Precision
Recall
F1-Score
IoU
Orthorectified
mag1c [12]
N/A ≥ 900
41.67 44.91
15.69 15.69
22.80 23.25
4.76 4.84
Orthorectified
U-Net
N/A ≥ 900
79.23 82.77
19.91 32.52
31.82 46.69
18.47 30.80
Unorthorectified
U-Net
N/A ≥ 900
88.41 89.08
19.85 33.01
32.42 48.16
16.91 26.17
Table 3: Semantic segmentation performance. Precision, recall, F1-score, and IoU are reported for mag1c [12] and U-Net, trained/evaluated on orthorectified and unorthorectified data. Best result per column within each block is highlighted in bold.
RGB & ground truth
U-Net
RGB & ground truth
Mag1c
(a) Orthorectified
U-Net
Mag1c
(b) Unorthorectified
Figure 3: Visualisation of semantic segmentation results on 2 example tiles from the orthorectified (Figure 3a) and the unorthorectified (Figure 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 U-Net; segmentation predictions from mag1c [12].
ble 4): FP16 conversion halves the model size (3.30 to 1.70 MB) with a maximum per-pixel output deviation below 0.3%, while INT8 quantisation reduces size further (to 0.84 MB, a 4× reduction) at the cost of up to ∼12% maximum output deviation. We consider FP16 the more robust operating point for onboard deployment given its negligible impact on outputs, while INT8’s larger deviation warrants further validation before deployment on hardware with stricter memory budgets.
3rd Conference on AI in and for Space (SPAICE 2026) © 2026 M. Chen et al., licenced via CC BY 4.0.
4
Discussion
In this work, we present UnorthoDOS, a dataset and training approach for methane plume detection directly on unorthorectified satellite imagery. We show that models trained on orthorectified and unorthorectified data achieve comparable performance, demonstrating that orthorectification can be bypassed: a critical advantage for real-time detection on resource-constrained satellites. We further
PyTorch File size (MB) Max relative difference (%)
32-bit ONNX
16-bit ONNX
8-bit Quant
3.30
3.25
1.70
0.84
0 0
1.66×10−4 2.19×10−4
0.30 0.25
11.93 10.70
Orthorectified Unorthorectified
Table 4: File sizes in MB, and the maximum relative difference in the outputs of the original PyTorch models trained on the orthorectifed and unorthorectified datasets and ONNX models compressed to varying precisions.
demonstrate onboard deployment feasibility via compression, halving model size with <1% output deviation. A key limitation is reduced sensitivity to weak methane plumes, mitigable by training on larger datasets as hyperspectral satellite deployments are increasing.
5
References
2.
Thompson, D. et al. Real-time remote detection and measurement for airborne imaging spectroscopy: a case study with methane. Atmospheric Measurement Techniques 8, 4383–4397 (2015).
12.
Foote, M. D. et al. Fast and accurate retrieval of methane concentration from imaging spectrometer data using sparsity prior. IEEE Transactions on Geoscience and Remote Sensing 58, 6480–6492 (2020).
13.
Meoni, G. et al. 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).
14.
Herec, J., Růžička, V. & Pitoňák, R. Optimizing Methane Detection On Board Satellites: Speed, Accuracy, and Low-Power Solutions for Resource-Constrained Hardware in 2025 European Data Handling & Data Processing Conference (EDHPC) (2025), 1–9.
15.
Růžička, V. et al. Semantic segmentation of methane plumes with hyperspectral machine learning models. Scientific Reports 13, 19999. issn: 2045-2322. doi:10.1038/s41598- 023- 44918- 6. https: //doi.org/10.1038/s41598-023-44918-6 (Nov. 2023).
16.
Růžička, V. & Markham, A. HyperspectralViTs: General Hyperspectral Models for On-Board Remote Sensing. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2025).
17.
Chen, M. et al. Towards Methane Detection Onboard Satellites. arXiv preprint arXiv:2509.00626 (2025).
18.
Green, R. EMIT L1B At-Sensor Calibrated Radiance and Geolocation Data 60 m V001 2022. doi:10 . 5067 / EMIT / EMITL1BRAD . 001.
Acknowledgments
This work has been enabled by Frontier Development Lab Earth Systems Lab (https://eslab.ai/) a public/private partnership between the European Space Agency (ESA), Trillium Technologies, the University of Oxford and leaders in commercial AI supported by Google Cloud, Scan Computers, Nvidia Corporation and Pasteur Labs.
1.
11.
Myhre, G. et al. in Climate Change 2013 - The Physical Science Basis 659–740 (Cambridge University Press, 2014). doi:10 . 1017 / CBO9781107415324 . 018. https : / / hal . science / hal 03323288. Jacob, D. J. et al. Satellite observations of atmospheric methane and their value for quantifying methane emissions. Atmospheric Chemistry and Physics 16, 14371–14396 (2016).
https : / / www . earthdata . nasa . gov / data / catalog / lpcloud-emitl1brad-001.
Green, R. et al. EMIT L2B Estimated Methane Plume Complexes 60 m V001 2023. doi:10.5067/EMIT/EMITL2BCH4PLM.001. https:
Barretta, D. et al. Toward Real-Time Earth System Intelligence Through Advancements in Onboard AI Processing: A review. IEEE Geoscience and Remote Sensing Magazine, 2–52. doi:10.1109/MGRS. 2026.3665978 (2026).
19.
4.
Rijlaarsdam, D. et al. The next era for earth observation spacecraft: An overview of cognisat-6. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 18, 2450–2463 (2024).
20.
Chen, M. et al. UnorthoDOS-Unorthorectified Dataset for On board Satellite methane detection. doi:10 . 57967 / hf / 6778. https : //huggingface.co/datasets/SpaceML/UnorthoDOS (2025).
5.
Furano, G. et al. Towards the use of artificial intelligence on the edge in space systems: Challenges and opportunities. IEEE Aerospace and Electronic Systems Magazine 35, 44–56 (2020).
21.
CEOS - Committee on Earth Observation Satellites. Greenhouse Gas (GHG) Emissions Monitoring: Point Source vs. Wide Area Mappers Aug. 2024. https://ceos.org/news/ghg-mappers/.
6.
Giuffrida, G. et al. The Φ-Sat-1 mission: The first on-board deep neural network demonstrator for satellite earth observation. IEEE Transactions on Geoscience and Remote Sensing 60, 1–14 (2021).
22.
7.
Marin, A. et al. Phi-Sat-2: Onboard AI apps for earth observation. Proc. Space Artif. Intell 6 (2021).
Ronneberger, O., Fischer, P. & Brox, T. U-net: Convolutional networks for biomedical image segmentation in International Conference on Medical image computing and computer-assisted intervention (2015), 234–241.
23.
8.
Ghasemi, N., Justo, J. A., Celesti, M., Despoisse, L. & Nieke, J. Onboard processing of hyperspectral imagery: Deep learning advancements, methodologies, challenges, and emerging trends. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 18, 4780–4790 (2025).
Sudre, C. H., Li, W., Vercauteren, T., Ourselin, S. & Jorge Cardoso, M. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations in International Workshop on Deep Learning in Medical Image Analysis (2017), 240–248.
24.
Kingma, D. P. & Ba, J. Adam: A Method for Stochastic Optimization in International Conference on Learning Representations (ICLR) (2015). https://arxiv.org/abs/1412.6980.
25.
ONNX Runtime developers. ONNX Runtime https : / / onnxruntime.ai/. 2021.
3.
9.
Marini, L. Semi-supervised learning for marine anomaly detection on board satellites. arXiv preprint arXiv:2504.03705 (2024).
10.
Manolakis, D., Truslow, E., Pieper, M., Cooley, T. & Brueggeman, M. Detection algorithms in hyperspectral imaging systems: An overview of practical algorithms. IEEE Signal Processing Magazine 31, 24–33 (2013).
3rd Conference on AI in and for Space (SPAICE 2026) © 2026 M. Chen et al., licenced via CC BY 4.0.
/ / www . earthdata . nasa . gov / data / catalog / lpcloud emitl2bch4plm-001.