Technical Summary: Towards Methane Detection On Board Satellites

Problem Statement

Methane is a potent greenhouse gas, with a global warming potential approximately 84 times that of carbon dioxide over a 20-year period. While hyperspectral remote sensing offers a unique opportunity to detect "super-emitters" from space, current detection pipelines face significant computational bottlenecks for on-board satellite deployment. Conventional methods typically require two resource-intensive preprocessing steps:

1. Orthorectification: Correcting geometric distortions caused by sensor viewing angles, terrain variations, and Earth curvature.
2. Matched Filtering: Generating methane enhancement products by comparing image data against predefined spectral signatures to strengthen weak plume signals.

These steps are computationally expensive, making real-time execution on resource-constrained spacecraft difficult. Furthermore, matched filters are prone to high false-positive rates. While recent deep learning approaches have improved segmentation, they remain reliant on orthorectified datasets, which are costly to generate in real-time on board.

Methodology

The authors propose a novel approach that bypasses the need for on-board orthorectification by utilizing UnorthoDOS (Unorthorectified Dataset for On Board Satellite methane detection).

Dataset Construction

• Source: Data is derived from the Earth Surface Mineral Dust Source Investigation (EMIT) sensor on the International Space Station (ISS), specifically L1B hyperspectral imagery and L2B methane plume complexes (ground truth).
• Spectral Selection: A subset of 86 bands (1573–1699 nm and 2004–2478 nm) was selected to capture methane absorption spectra alongside non-absorption and RGB bands.
• UnorthoDOS Generation: Instead of processing raw data into orthorectified images, the authors generated an unorthorectified dataset by applying an inverse geometric mapping ($\tau^{-1}$) to the orthorectified ground truth annotations. This process reconstructs the annotations in the original off-nadir (unorthorectified) pixel coordinates using nearest-neighbor interpolation to handle non-surjective mappings.
• Benchmark: A corresponding orthorectified dataset was generated using the same L1B images and original L2B annotations for comparative analysis.
• Data Splitting: The data was tiled into 128 × 128 pixels. A holdout test set (5% of images) was reserved, with the remainder split into training (80%) and validation (15%) sets. Spatial jittering was applied to augment the limited number of labeled plumes.

Experimental Setup

The authors implemented a simplified "tip and cue" paradigm:

• Tip Satellite: Performs image classification to determine the presence of a plume.
• Cue Satellite: Performs semantic segmentation to map the plume's extent.
• Model Architecture: A UNet architecture was used for both tasks.
• Training: Models were trained independently on orthorectified and unorthorectified datasets using Dice loss, the Adam optimizer, a learning rate of $1e^{-4}$, and a batch size of 32 for 100 epochs.
• Baselines: Performance was compared against the mag1c matched filter baseline.

Key Contributions

1. UnorthoDOS Dataset: The release of the first ML-ready dataset of unorthorectified hyperspectral images specifically designed for on-board methane detection, alongside a corresponding orthorectified benchmark.
2. Preprocessing Bypass: Demonstration that ML models can be trained effectively on unorthorectified data, eliminating the need for computationally heavy orthorectification on the satellite.
3. Open Resources: Public release of model checkpoints, the datasets (via Hugging Face), and code (via GitHub) to facilitate further research in space-based ML.

Results

The study evaluated performance using Precision, Recall, F1-Score, Accuracy (for classification), and Intersection over Union (IoU) (for segmentation).

• vs. Matched Filter (mag1c): The UNet models significantly outperformed the mag1c baseline. For weak plumes, IoU improved by 288.03%, and for strong plumes (≥900 ppm m), IoU improved by 536.36%. The UNet provided more coherent spatial delineation, whereas mag1c produced fragmented outputs with higher false-positive rates due to misclassifying terrestrial features.
• Orthorectified vs. Unorthorectified:
  • Classification: Models trained on orthorectified data showed slightly superior results (e.g., 94.60% recall vs. 71.61% for strong plumes), potentially due to better spatial alignment or image quality.
  • Segmentation: Performance was comparable between the two datasets. Interestingly, the model trained on unorthorectified data achieved slightly better segmentation metrics in some areas, though the orthorectified model yielded a higher IoU for strong plumes (30.80% vs. 26.17%).
  • Weak Plumes: Absolute performance for weak plumes remained limited (IoU between 16% and 19%) across both datasets, indicating a sensitivity ceiling.
• Strong Plumes: Performance improved substantially for plumes with concentrations ≥900 ppm m, with significant gains in recall and IoU for both datasets.

Significance and Claims

The paper claims that methane plume detection is feasible without orthorectification, offering a critical advantage for real-time detection in resource-constrained satellite scenarios. By demonstrating that ML models trained on unorthorectified data achieve performance comparable to those trained on orthorectified data, the authors suggest a pathway to deploy rapid, on-board early detection systems that reduce downlink costs and support faster mitigation responses.

The authors modestly acknowledge a key limitation: the reduced sensitivity to weak methane plumes. They posit that this limitation could be mitigated by access to larger annotated hyperspectral datasets, which will become available as the number of satellite missions with dedicated sensors increases. The work establishes a foundation for moving beyond traditional, computationally heavy preprocessing pipelines in favor of direct, end-to-end learning on raw satellite data.