Semantic segmentation of methane plumes with hyperspectral machine learning models.

Area of Science:

• Environmental Science
• Remote Sensing
• Machine Learning

Background:

• Methane is a potent greenhouse gas, and its mitigation is crucial for climate change prevention.
• Fossil fuel industry point-sources offer significant methane mitigation potential.
• Current methane plume detection methods have high false positive rates and require manual input.

Purpose of the Study:

• To address the lack of large, annotated datasets for methane plume detection.
• To develop and benchmark sensor-agnostic machine learning models for methane plume identification.
• To improve the accuracy and efficiency of detecting methane emissions from remote sensing data.

Main Methods:

• Publicly released a machine learning-ready dataset with manually annotated methane plumes.
• Utilized hyperspectral data from AVIRIS-NG and simulated multispectral WorldView-3 data.
• Proposed sensor-agnostic machine learning architectures (HyperSTARCOP) using methane enhancement products.

Main Results:

• HyperSTARCOP model achieved over 25% higher F1 score and reduced false positives by over 41.83% compared to baseline methods.
• Demonstrated zero-shot generalization on EMIT hyperspectral instrument data.
• Achieved a 40% gain in F1 score on an annotated subset of EMIT images.

Conclusions:

• The developed dataset and HyperSTARCOP model offer a significant advancement in methane plume detection.
• Sensor-agnostic models show promise for cross-sensor methane emission monitoring.
• Improved detection capabilities can enhance climate change mitigation strategies.