Few-Shot Contrail Segmentation in Remote Sensing Imagery With Loss Function in Hough Space

Abstract

Condensation trails, or contrails, are line-shaped clouds that are produced by an aircraft engine exhaust. These contrails often impact climate significantly due to their potential warming effect. Identification of contrail formation through satellite images has been an ongoing research challenge. Traditional computer vision techniques struggle with varying imagery conditions, and supervised machine learning approaches often require a large amount of hand-labeled images. This study researches few-shot transfer learning and provides an innovative approach for contrail segmentation with a few labeled images. The methodology leverages backbone segmentation models, which are pretrained on existing image datasets and fine-tuned using an augmented contrail-specific dataset. We also introduce a new loss function, SR loss, which enhances contrail line detection by incorporating Hough transformation in model training. This transformation improves performance over generic image segmentation loss functions. The openly shared few-shot learning library, contrail-seg, has demonstrated that few-shot learning can be effectively applied to contrail segmentation with the new loss function.