Source-linked AI summary
Machine Vision for Natural Gas Methane Emissions Detection Using an Infrared Camera
Jingfan Wang, Lyne P. Tchapmi, Arvind P. Ravikumara, Mike McGuire, Clay S. Bell, Daniel Zimmerle, Silvio Savarese, Adam R. Brandt
TL;DR
Manual OGI surveys face high labor costs and cannot provide leak-detection results without operator judgment. This paper explores computer-vision approaches for automated methane-leak detection and demonstrates GasNet successfully using the GasVid dataset.
Problem
Manual OGI surveys have high labor costs, while continuous monitoring with infrared cameras is infeasible.
Method
The paper explores computer-vision approaches to determine whether a methane leak is present.
Results
GasNet is demonstrated to be successful using the GasVid dataset.
Takeaways & Limitations
The machine-vision algorithm provides a demonstrated approach for methane-emission detection.
Takeaways & Limitations
The authors cannot draw a conclusion that the algorithm’s performance covers the diversity of leaks observed in the real world.
Abstract
from arXiv · showhide
It is crucial to reduce natural gas methane emissions, which can potentially offset the climate benefits of replacing coal with gas. Optical gas imaging (OGI) is a widely-used method to detect methane leaks, but is labor-intensive and cannot provide leak detection results without operators' judgment. In this paper, we develop a computer vision approach to OGI-based leak detection using convolutional neural networks (CNN) trained on methane leak images to enable automatic detection. First, we collect ~1 M frames of labeled video of methane leaks from different leaking equipment for building CNN model, covering a wide range of leak sizes (5.3-2051.6 gCH4/h) and imaging distances (4.6-15.6 m). Second, we examine different background subtraction methods to extract the methane plume in the foreground. Third, we then test three CNN model variants, collectively called GasNet, to detect plumes in videos taken at other pieces of leaking equipment. We assess the ability of GasNet to perform leak detection by comparing it to a baseline method that uses optical-flow based change detection algorithm. We explore the sensitivity of results to the CNN structure, with a moderate-complexity variant performing best across distances. We find that the detection accuracy can reach as high as 99%, the overall detection accuracy can exceed 95% for a case across all leak sizes and imaging distances. Binary detection accuracy exceeds 97% for large leaks (~710 gCH4/h) imaged closely (~5-7 m). At closer imaging distances (~5-10 m), CNN-based models have greater than 94% accuracy across all leak sizes. At farthest distances (~13-16 m), performance degrades rapidly, but it can achieve above 95% accuracy to detect large leaks (>950 gCH4/h). The GasNet-based computer vision approach could be deployed in OGI surveys to allow automatic vigilance of methane leak detection with high detection accuracy in the real world.
1. Introduction
Natural gas can provide climate benefits relative to coal, but methane leakage may offset them, creating a need for more reliable leak detection. This paper addresses limitations of manual OGI by developing CNN-based automatic detection for infrared methane-leak images.
- Motivation: Methane’s high global warming potential can undermine the climate benefits of replacing coal with natural gas.Methane is reported as 36 times more potent per kilogram than CO2 over 100 years.
- Motivation: The largest 5% of leaks account for approximately 50% of total emissions, making effective leak identification important.
- Challenges: Manual OGI surveys are labor-intensive, while continuous infrared-camera monitoring is infeasible.
- Challenges: Infrared cameras cannot provide real-time leak-detection results without operator judgment, and survey quality varies between operators.
- Approach: The paper develops CNN-based computer vision to determine whether infrared images contain methane leaks.
- Approach: The workflow builds labeled leak videos, tests background subtraction, develops GasNet, and evaluates detection across distances, leak sizes, and environmental conditions against optical-flow algorithms.
2. Related work
Prior automated image-analysis research includes change detection, texture analysis, color modeling, and machine learning, but methane detection remains comparatively underdeveloped. This paper applies deep learning to infrared methane-leak imagery rather than relying on hand-crafted features.
- Methane versus smoke: Methane plumes resemble smoke in dispersion and irregular motion but require infrared imaging, whereas smoke is captured in RGB.
- Existing image-analysis methods: Smoke-detection research commonly uses color modeling, change detection, texture analysis, and machine-learning models.
- Existing image-analysis methods: Change detection separates moving leak regions from static backgrounds by subtracting an estimated background and thresholding the residual.
- Research gap: Deep learning has been applied broadly in energy and environmental applications, but research on methane-emission detection remains limited.
- Paper approach: The paper uses CNNs to analyze methane leaks recorded by a FLIR GF-320 infrared camera instead of using hand-crafted features.
3. Datasets-GasVid
GasVid is a controlled-release, labeled methane-leak video dataset spanning multiple leak sources, sizes, orientations, environmental conditions, and imaging distances. The dataset separates training and testing equipment but uses otherwise simplified recording conditions and excludes several real-world disturbances.
- Dataset construction: GasVid contains labeled methane-leak videos from multiple leakage sources covering a wide range of leak sizes.
- Dataset construction: Controlled-release experiments at METEC recorded separator and tank leaks using a FLIR GF-320 infrared camera.
- Data limitations: Tank videos were unusable because leak orientation and camera location were poor.
- Data partitioning: Separator-2 data are used for CNN training and validation, while separator-1 data are reserved exclusively for testing.
- Coverage: The videos span five imaging distances from 4.6 m to 15.6 m and eight leak classes ranging from 5.3 to 2051.6 gCH4/h.
- Data limitations: The study uses simplified conditions: a tripod prevents camera movement, and videos exclude moving vegetation, vapor, and steam.
- Data limitations: Plumes become harder to observe with increasing distance, while uncontrolled wind affects their appearance.
4. Method
The method combines background subtraction, image normalization, and CNN-based binary classification to automatically detect methane plumes. GasNet variants are evaluated against a non-CNN baseline across leak classes and imaging distances.
- Workflow: The workflow introduces three background-subtraction methods, image normalization, three CNN variants, and a non-CNN baseline for comparison.Processed frames are fed into the trained CNN, and prediction accuracy is calculated from correct identifications.
- Experimental design: Each binary case classifies images as non-leak or one of seven leak classes at each of five imaging distances, producing 35 cases.The experiments use independent training runs under different distance and leak-size conditions.
- Background subtraction: Background subtraction extracts foreground plume features, but real methods leave some non-plume features in the image.The paper compares fixed, moving-average, and Mixture-of-Gaussians methods with a no-subtraction baseline.
- Background subtraction: Moving-average backgrounds use the median of the previous 210 images, equivalent to a lagged 14-second video window.This smooths plume variations over multiple seconds before subtraction.
- Background subtraction: For stable backgrounds, all three methods isolate the plume, whereas moving-average and MOG subtraction remove a moving cloud treated as background.Fixed subtraction retains the cloud because its static background does not account for recent cloud movement.
5. Results and discussion
GasNet performance depends strongly on background subtraction, imaging distance, leak size, architecture, and training aggregation. Moving-average subtraction and the moderate-complexity GasNet-2 generally provide the strongest results, while accuracy declines for small leaks at long distances.
- Background subtraction: Accuracy improves with larger leak classes but degrades toward near-random performance as imaging distance increases and leak size decreases.
- Background subtraction: 93% accuracy is exceeded at 4.6 m and 6.9 m using fixed-average subtraction, but accuracy falls at 9.8 m and 12.6 m for smaller leak classes.
- Background subtraction: Moving-average subtraction exceeds 97% accuracy at 4.6 m, exceeds 98% for leak classes 2–7 at 6.9 m, and reaches 99%.
- Background subtraction: MOG subtraction performs better than fixed-average subtraction but worse than moving-average subtraction, with gaps of 3.9%, 3.3%, and 6.4% at 9.8, 12.6, and 15.6 m.
- Architecture: GasNet-2 outperforms GasNet-1 and GasNet-3, while CNN-based analysis performs better than the optical-flow baseline.
- Distance and aggregation: Binary detection accuracy exceeds 94% across leak sizes at approximately 5–10 m, whereas accuracy decreases sharply for the smallest leaks at the longest distances.
6. Concluding remarks
The paper demonstrates a CNN-based GasNet approach for automated methane-leak detection using background subtraction and the GasVid dataset. The authors identify practical deployment opportunities while noting that generalization and accuracy remain dependent on leak diversity and imaging conditions.
- Background subtraction combined with GasNet is presented as an appropriate approach for leak detection that can be trained for complex real-world environments.
- Accuracy results from this automated technology cannot be directly compared with prior OGI results obtained by human operators.
- The CNN-based machine-vision algorithm is demonstrated for methane-emission detection using the GasVid dataset.
- Future dataset expansion is intended to represent real-world leak diversity and improve GasNet generalizability.
- Algorithm accuracy depends on camera performance, imaging distance, temperature contrast, and background movement.
- Automated OGI could support automatic vigilance from maintenance vehicles or perimeter security cameras, reducing manual detection labor and accelerating leak detection and repair.