Source-linked AI summary
Methane Detection On Board Satellites from Unorthorectified Imagery
Luca Marini, Maggie Chen, Hala Lamdouar, Laura Martínez-Ferrer, Dr C. P. Bridges, Giacomo Acciarini
TL;DR
Timely methane detection is important, but conventional orthorectification and matched-filter processing are poorly suited to onboard execution. The paper introduces UnorthoDOS for training directly on unorthorectified hyperspectral imagery, achieving comparable segmentation performance while supporting compact deployment.
Problem
Conventional methane detection relies on orthorectification and matched-filter products that are poorly suited to onboard execution.
Method
UnorthoDOS trains machine-learning models directly on unorthorectified hyperspectral imagery, using EMIT-derived datasets and U-Net models.
Results
16.91% IoU on all plumes for the unorthorectified U-Net approaches 18.47% for the orthorectified U-Net, while both exceed mag1c’s 4.76%.
Takeaways & Limitations
Unorthorectified inference can bypass orthorectification, and FP16 compression halves model size with below 0.3% maximum per-pixel output deviation.
Takeaways & Limitations
Sensitivity is reduced for weak methane plumes, which the authors state may be mitigated by training on larger datasets.
Abstract
from arXiv · showhide
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 U-Net 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’s high warming impact and the detectability of major point sources motivate rapid satellite-based detection. Existing workflows depend on ground-oriented orthorectification and computationally costly matched filters, motivating UnorthoDOS for unorthorectified imagery.
- Methane has approximately 84 times carbon dioxide’s 20-year global warming potential, and major oil-and-gas super-emitters are detectable from space.
- Onboard AI combined with hyperspectral imaging could reduce methane-detection latency and downlink requirements.
- Orthorectification corrects viewing-geometry distortions, while matched filters enhance weak plume signals using predefined spectral signatures.
- Orthorectification is intended for ground processing, and traditional matched filters are computationally costly and prone to false positives.
- UnorthoDOS introduces an unorthorectified hyperspectral dataset and approach that better reflect onboard detection conditions.
2 Methodology
The methodology constructs paired orthorectified and unorthorectified EMIT datasets, then trains and evaluates U-Net models for plume classification, segmentation, and compressed onboard deployment.
- Data source: The corpus contains 1,574 annotated EMIT methane plumes, using 86 selected bands spanning 1573-1699 nm and 2004-2478 nm plus three RGB bands.
- Dataset generation: Unorthorectified imagery is approximated by applying the inverse geometric mapping to orthorectified imagery and filling missing values with nearest-neighbor interpolation.
- Dataset generation: Hyperspectral images and plume annotations are tiled into 128 × 128 pixels to create the UnorthoDOS ML-ready dataset.
- Experimental design: The orthorectified benchmark uses the same EMIT scenes and plumes as UnorthoDOS, isolating the effect of orthorectification.
- Detection pipeline: A tip-and-cue setup uses binary classification to flag candidate plumes and semantic segmentation to localize them precisely.
- Evaluation: Scene-level splits use 80% training, 15% validation, and 5% holdout testing, with identical splits across geometric variants.
- Model training: Two 86-channel U-Net models are trained separately on orthorectified and unorthorectified data to predict binary per-pixel plume masks.
- Evaluation: Performance is evaluated using classification and segmentation metrics, including precision, recall, F1-score, accuracy, and IoU.
3 Results
U-Net models substantially outperform mag1c for semantic segmentation, with unorthorectified training approaching orthorectified performance. FP16 compression also reduces model size with minimal output deviation.
- Segmentation performance: 18.47% vs. 4.76% IoU: the orthorectified U-Net outperforms mag1c on all plumes.The unorthorectified U-Net achieves 16.91% IoU against the same 4.76% mag1c baseline.
- Segmentation performance: 30.80% vs. 4.84% IoU: the orthorectified U-Net outperforms mag1c on strong plumes.The unorthorectified U-Net achieves 26.17% IoU against 4.84% for mag1c.
- Plume strength: 18.47% to 30.80% and 16.91% to 26.17%: segmentation IoU rises for orthorectified and unorthorectified models on strong plumes.Classification recall also rises from 56.21% to 83.33% and from 48.89% to 71.61%, respectively.
- Geometric variants: 16.91% vs. 18.47% IoU: unorthorectified and orthorectified U-Nets achieve comparable all-plume segmentation performance.Classification performance shows a wider unorthorectified gap, with accuracy 5.3 points lower on all plumes and 6.8 points lower on strong plumes.
- Onboard deployment: 3.30 to 1.70 MB: FP16 conversion halves model size with maximum per-pixel output deviation below 0.3%.INT8 reduces size to 0.84 MB, a 4× reduction, but reaches up to ∼12% maximum output deviation.
4 Discussion
UnorthoDOS supports methane plume detection directly from unorthorectified imagery, while FP16 compression supports smaller onboard models with minimal output deviation. Reduced sensitivity to weak plumes remains a limitation.
- Discussion: Comparable performance between orthorectified and unorthorectified models demonstrates that orthorectification can be bypassed.This supports real-time detection on resource-constrained satellites.
- Discussion: FP16 compression halves model size with under 1% output deviation, demonstrating onboard deployment feasibility.The paper identifies compression as a route toward onboard execution.
- Limitations: Reduced sensitivity to weak methane plumes limits the approach, with larger training datasets proposed as a mitigation.The limitation is framed in the context of increasing hyperspectral satellite deployments.