17:15 〜 19:15
[ACG41-P24] Detecting contrails from satellite images to assess their climate impact
キーワード：飛行機雲、衛星観測、放射収支、radiation budget
Contrails, or condensation trails, are line-shaped clouds formed by aircraft as their exhaust gases cool rapidly in the atmosphere. Those that form in ice-supersaturated regions (ISSR) can persist for hours and evolve into contrail cirrus, which has a net positive radiative forcing (RF), making it the second-largest contributor to aviation-induced climate change after CO2 emissions. Contrail cirrus alters the Earth's energy balance by trapping outgoing longwave radiation while reflecting some incoming solar radiation. The overall effect is a net warming, contributing significantly to climate change. According to the IPCC AR6 (2023), contrail cirrus is responsible for approximately 5% of the total RF from human activities. While its global impact is relatively small, regional effects are significant, with predicted RF increases over Asia (Teoh et al., 2023). Understanding the regional variations in contrail RF is essential for assessing the climate impact of aviation. Given the rapid growth of air traffic in Asia, it is crucial to analyze the long-term impact of contrails in this region and compare it with other high-traffic zones like the North Atlantic and the Americas.
This study aims to statistically analyze contrails detected in satellite images and compare the RF of contrails over Asia, North Atlantic, and the Americas. The study covers the period from January 2016 to December 2021, including the COVID-19 pandemic. The workflow consists of detecting contrails from satellite images using a deep learning model, validating the model against labeled datasets, comparing contrail RF over different regions, and assessing regional trends in contrail formation.
The detection model was developed using the OpenContrails dataset (Ng et al., 2023), which is based on GOES-16 Advanced Baseline Imager (ABI) data. Different spectral bands and time steps were tested to optimize detection accuracy. The model uses NeXUnet, a custom neural network architecture, to segment contrails from satellite images and generate binary contrail masks. The dataset includes images from GOES-16, a geostationary satellite positioned at 75.2°W with an altitude of 35,786 km and contains nine infrared channels and human-labeled ground truth data, providing detailed spectral information useful for detecting contrails and distinguishing them from other cloud types, ensuring reliable contrail identification and classification.
The model was able to correctly identify more than half of the contrails, demonstrating the feasibility of developing a detection model. Although the model successfully detected contrails, several issues were observed: non-contrail pixels with darker tones were sometimes misclassified as contrails, and very few pixels were classified as contrails overall. To improve detection accuracy, additional training data, particularly those without contrails, are needed. Including more spectral bands in the model and integrating flight path data and ice saturation levels may also enhance performance.
This study developed a contrail detection model and demonstrated its potential in identifying contrails and estimating their climate impact. By improving detection methods and analyzing regional RF variations, this research aims to contribute to a better understanding of aviation-induced climate effects.
This study aims to statistically analyze contrails detected in satellite images and compare the RF of contrails over Asia, North Atlantic, and the Americas. The study covers the period from January 2016 to December 2021, including the COVID-19 pandemic. The workflow consists of detecting contrails from satellite images using a deep learning model, validating the model against labeled datasets, comparing contrail RF over different regions, and assessing regional trends in contrail formation.
The detection model was developed using the OpenContrails dataset (Ng et al., 2023), which is based on GOES-16 Advanced Baseline Imager (ABI) data. Different spectral bands and time steps were tested to optimize detection accuracy. The model uses NeXUnet, a custom neural network architecture, to segment contrails from satellite images and generate binary contrail masks. The dataset includes images from GOES-16, a geostationary satellite positioned at 75.2°W with an altitude of 35,786 km and contains nine infrared channels and human-labeled ground truth data, providing detailed spectral information useful for detecting contrails and distinguishing them from other cloud types, ensuring reliable contrail identification and classification.
The model was able to correctly identify more than half of the contrails, demonstrating the feasibility of developing a detection model. Although the model successfully detected contrails, several issues were observed: non-contrail pixels with darker tones were sometimes misclassified as contrails, and very few pixels were classified as contrails overall. To improve detection accuracy, additional training data, particularly those without contrails, are needed. Including more spectral bands in the model and integrating flight path data and ice saturation levels may also enhance performance.
This study developed a contrail detection model and demonstrated its potential in identifying contrails and estimating their climate impact. By improving detection methods and analyzing regional RF variations, this research aims to contribute to a better understanding of aviation-induced climate effects.