Automatic detection of methane emissions in multispectral satellite imagery using a vision transformer

Methane, a potent greenhouse gas, is a significant contributor to global warming. While its shorter atmospheric lifetime compared to CO₂ makes reducing methane emissions crucial for short-term climate action, current atmospheric levels are steadily increasing. Precise identification and quantification of methane sources are vital for prioritizing mitigation efforts, but existing methods face significant challenges. Bottom-up inventories often underestimate emissions, and current measurement approaches struggle with coverage, resolution, and accuracy trade-offs. Ground-based and airborne measurements have limited coverage, while hyperspectral satellites offer high spectral resolution but compromise on spatial resolution or coverage. Multispectral satellites like Sentinel-2 offer high spatial and temporal resolution with global coverage but lack sufficient spectral information for accurate methane detection. This paper addresses this challenge by developing a novel deep learning approach tailored to multispectral satellite data, aiming to automatically identify methane signatures and improve detection capabilities beyond the state-of-the-art.

Existing methane detection methods rely on identifying methane absorption in the short-wave infrared (SWIR) region of backscattered sunlight. Hyperspectral satellites provide high SWIR spectral resolution for precise methane concentration determination but suffer from limited coverage or resolution. Machine learning techniques have been applied to improve hyperspectral data analysis. However, multispectral satellite data, while providing high spatial and temporal resolution and global coverage, has limitations due to reduced spectral information. Previous work with multispectral data, like Sentinel-2, has only enabled the detection of large emissions (2-3 tons/h or greater). This study builds upon these previous attempts, particularly those utilizing deep learning for methane plume identification from hyperspectral data and multispectral data analysis methods using band ratios, to develop a more sensitive and accurate approach.

This research utilized a deep learning architecture specifically designed for multispectral satellite data, aiming to automatically identify methane signatures while minimizing noise. Due to limited ground truth data for methane detection in multispectral images, the researchers employed a synthetic data generation approach. A large database of Sentinel-2 tiles, sampled at consecutive times with minimal cloud cover, was compiled from various regions representing different climates, topographies, and land uses. These tiles were divided into 2.5 x 2.5 km² scenes. Synthetic methane plumes, generated using a simplified Gaussian plume model (instead of computationally expensive WRF-LES simulations), were embedded into about half of these scenes using the Beer-Lambert law, while the other half served as negative examples. These synthetic plumes varied in emission rates and wind velocities and incorporated auto-correlated atmospheric noise to mimic real-world conditions. The resulting dataset contained approximately 1,650,000 unique Sentinel-2 samples. A novel encoder-decoder architecture was trained. The encoder utilized a Vision Transformer (ViT) to capture long-range dependencies within the plume structures, while the decoder, a U-Net architecture, facilitated the reconstruction of the plume's spatial extent. The input to the model comprised ten spectral bands from two consecutive time steps. The model learned to identify methane leaks by classifying pixels corresponding to the embedded synthetic plumes. The dataset was split into training, validation, and testing sets (75%, 10%, and 15%, respectively), with distinct geographic locations for each set to enhance model generalization. The model was trained for 10 epochs using the Adam optimizer with a learning rate that decreased dynamically. Model performance was evaluated using the F1 score (harmonic mean of precision and recall). The Multi-band multi-pass method (MBMP) was used as a benchmark for comparison. For evaluating real methane plumes, the researchers used the Methane Plumes from Airborne Surveys open-source dataset published by Carbon Mapper, which includes plumes detected during airborne surveys across various U.S locations. The deep learning model was applied to Sentinel-2 images close in time to the airborne detections. The model's performance was assessed by comparing its detections to the cataloged plumes, considering factors like plume extent and emission rates. A controlled release experiment provided further independent validation. The code is available online.

The developed deep learning model significantly improved methane detection capabilities compared to state-of-the-art methods. On synthetic data, the model reliably detected methane plumes down to a signal-to-noise ratio (SNR) of about 5%, an order of magnitude better than the MBMP method. The model achieved an extremely low false positive rate, a crucial aspect for automated detection. When applied to real methane plumes from airborne surveys (Carbon Mapper dataset), the model effectively detected plumes down to 200-300 kg/h, with detection rates close to the average persistence of leaks in the catalog. Smaller plumes (around 60 kg/h) were also detected. Residual false positives were primarily attributed to clouds, rivers, and soil moisture changes. The model's performance was more sensitive to plume extent than to emission rate. Independent validation using controlled release experiments further demonstrated the model's ability to detect large methane emitters with minimal false positives, even successfully identifying a 1.1 ton/h leak previously missed by other methods. The study highlights that the deep learning approach using a ViT-based encoder significantly improves detection capabilities.

This study's findings significantly advance the capabilities of methane detection using multispectral satellite imagery. The deep learning model's ability to detect plumes an order of magnitude smaller than previously possible with multispectral data significantly reduces the existing trade-off between spatial and spectral resolution. The combination of high accuracy, low false positive rates, and global coverage makes this approach a powerful tool for monitoring methane emissions. The results suggest that publicly available, general-purpose multispectral satellites can be effectively leveraged for high-resolution methane monitoring, approaching the performance of dedicated hyperspectral constellations. This could contribute significantly to building global methane inventories at fine spatial and temporal scales, facilitating more effective mitigation strategies.

This research demonstrates the potential of deep learning, specifically using Vision Transformers, for accurate and automated detection of methane emissions from multispectral satellite data. The method achieves an order of magnitude improvement over state-of-the-art techniques, enabling detection down to plumes of 200-300 kg/h. The high accuracy, low false positive rate, and global coverage make it a valuable tool for methane monitoring and inventory building. Future work will focus on reducing residual false positives, incorporating auxiliary data, scaling up for global-scale detection, and further validation using controlled release experiments with smaller emission rates.

While the model demonstrates significant improvements, some limitations exist. The primary limitation arises from the reliance on synthetic data for training. Although the synthetic plumes incorporated realistic noise, some discrepancies might exist between the synthetic and real-world plume characteristics. Additionally, the temporal offset between satellite and airborne measurements in the real-world validation introduces some uncertainty. Finally, the remaining false positives, primarily due to clouds, rivers, and soil moisture variations, require further refinement.