Model Card

This Hugging Face repository contains models trained in the article "Optimizing Methane Detection On Board Satellites: Speed, Accuracy, and Low-Power Solutions for Resource-Constrained Hardware."

Paper: Optimizing Methane Detection On Board Satellites: Speed, Accuracy, and Low-Power Solutions for Resource-Constrained Hardware

Model Overview

The models here were trained using the code available at the following GitHub repository:

Training Code: HyperspectralViTs

The main project code, including filters benchmarking and demos, is available at the:

Project Code: Methane Filters Benchmark

Data and Products

The precomputed products used for training were created by code in the main project repository:

Products Creation Code: methane-filters-benchmark

Additionally, these precomputed products are hosted and accessible here:

Dataset Repository: STARCOP-fast-products

Sample Usage

You can try out our models and demos directly in Google Colab using the provided notebooks:

Models Demo: This notebook demonstrates model inference.
Products Creation and Benchmarking Demo: This notebook demonstrates generating products and measuring their runtime.

For local inference using the ONNX models, refer to the benchmark/onnx_inference_time.py script in the Project Code repository.

Citation

If you use these models in your research, please cite our article:

@misc{herec2025optimizingmethanedetectionboard,
      title={Optimizing Methane Detection On Board Satellites: Speed, Accuracy, and Low-Power Solutions for Resource-Constrained Hardware},
      author={Jonáš Herec and Vít Růžička and Rado Pitoňák},
      year={2025},
      eprint={2507.01472},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.01472},
}

Models performance

U-Net - CEM

ID	Recall	Precision	F1	F1 strong
A	0.441	0.317	0.369	0.500
B	0.701	0.158	0.258	0.550
C	0.531	0.299	0.382	0.610
D	0.536	0.218	0.310	0.551
E	0.564	0.182	0.275	0.469
AVG	55.47%	23.49%	31.90%	53.55%
STD	8.41%	6.30%	4.94%	4.85%

U-Net - ACE

ID	Recall	Precision	F1	F1 strong
A	0.468	0.202	0.282	0.460
B	0.480	0.288	0.360	0.537
C	0.413	0.253	0.314	0.461
D	0.550	0.194	0.287	0.510
E	0.500	0.162	0.245	0.442
AVG	48.22%	21.99%	29.77%	48.19%
STD	4.46%	4.50%	3.82%	3.57%

U-Net - MF

ID	Recall	Precision	F1	F1 strong
A	0.603	0.153	0.243	0.451
B	0.673	0.198	0.306	0.585
C	0.563	0.259	0.355	0.507
D	0.625	0.173	0.271	0.558
E	0.466	0.301	0.366	0.496
AVG	58.60%	21.68%	30.82%	51.94%
STD	6.97%	5.52%	4.73%	4.73%

U-Net - MAG1C-SAS

ID	Recall	Precision	F1	F1 strong
A	0.587	0.456	0.513	0.668
B	0.618	0.291	0.395	0.642
C	0.576	0.290	0.386	0.604
D	0.613	0.414	0.495	0.686
E	0.427	0.280	0.338	0.470
AVG	56.42%	34.62%	42.54%	61.40%
STD	7.04%	7.38%	6.73%	7.71%

U-Net - MAG1C (tile-wise)

ID	Recall	Precision	F1	F1 strong
A	0.643	0.218	0.325	0.599
B	0.732	0.288	0.413	0.692
C	0.613	0.362	0.455	0.659
D	0.669	0.242	0.355	0.633
E	0.640	0.366	0.466	0.684
AVG	65.94%	29.52%	40.28%	65.34%
STD	4.04%	6.05%	5.51%	3.42%

LinkNet - CEM

ID	Recall	Precision	F1	F1 strong
A	0.597	0.319	0.416	0.633
B	0.539	0.274	0.363	0.603
C	0.452	0.233	0.308	0.527
D	0.606	0.165	0.260	0.561
E	0.442	0.144	0.217	0.455
AVG	52.72%	22.70%	31.27%	55.56%
STD	6.96%	6.54%	7.09%	6.20%

LinkNet - MAG1C-SAS

ID	Recall	Precision	F1	F1 strong
A	0.566	0.324	0.412	0.612
B	0.505	0.515	0.510	0.613
C	0.381	0.422	0.400	0.507
D	0.590	0.383	0.464	0.660
E	0.513	0.378	0.435	0.627
AVG	51.10%	40.44%	44.42%	60.38%
STD	7.24%	6.35%	3.95%	5.14%