Ground-Based Contrail Detection by Means of Computer Vision Models: A Comparison Between Visible and Infrared Images

This paper deals with the problem of automatically detecting contrails, investigating the possible usage of infrared and visible images, through the construction of a labeled dataset from ground-based all-sky camera images. The images are preprocessed to remove the geometrical distortion introduced by the fisheye lens of the cameras and then fed to three different models, namely Roboflow Train 3.0, YOLOv8s, and YOLOv9c. The performances of the three approaches are then compared. The results show that the models trained on visible images outperform those trained on the infrared ones, with yolov9c being the best on both types of images, followed closely by yolov8s.