Deep Semantic Contrails Segmentation of GOES-16 Satellite Images: A Hyperparameter Exploration

This paper presents a comprehensive study on the optimization of hyperparameters for deep semantic segmentation models aimed at detecting contrails in GOES-16 satellite imagery. The environmental impact of aviation contrails has received considerable attention due to their potential contribution to climate change. Accurate contrail detection is essential for developing strategies to mitigate these impacts. Using the OpenContrails dataset [1] and advanced computer vision techniques, we performed a greedy hyperparameter search over different neural architectures, loss functions, and preprocessing methods. Our results indicate that using CoatNet as the backbone, coupled with the Unet++ architecture and dice loss as the optimization criterion, yields superior performance in contrail segmentation. In addition, incorporating data augmentation and resizing images to 512 pixels significantly improves model accuracy and generalization. The optimized model configurations demonstrate a promising approach for improving contrail segmentation, contributing to more accurate climate impact assessments and the development of sustainable aviation practices.