Strengthening Long-range Dependencies in Deep Convolutional Neural Networks for Contrail Detection in Satellite Imagery

Contrails, formed as aircraft pass through ice-supersaturated regions, contribute to aviation induced cloudiness and are known to have a significant warming effect on the planet, exceeding that of other non-CO₂ emissions in aviation. The potential for contrail avoidance and the resulting reduction in climate impact through flight rerouting underscores the importance of detecting and tracking contrails in satellite imagery and identifying sensitive regions. Given the extensive linear patterns of contrails observable in satellite imagery, we propose integrating non local attention blocks into U-Net, a convolutional network tailored for semantic segmentation, to strengthen long-distance dependencies. Experiments on a human-labeled contrail satellite imagery dataset show that incorporating the attention blocks yields superior performance compared to baseline models.