Exploring Models and Band Selection for Improved Contrail Detection with Deep Learning

Abstract

The consequences of climate change are becoming increasingly urgent, with contrails emerging as a potential contributing factor to this phenomenon. Consequently, there is an urgent need for precise techniques to detect them in satellite imagery. This study uses deep learning models and band selection to improve contrails detection in geostationary satellite imagery, using the Landsat-8 dataset with human-labeled contrails sourced from the GOES-16 Advanced Baseline Imager. By comparing different deep learning model methods such as DeepLabV3, U-Net, Fully Convolutional Network (FCN), Pyramid Scene Parsing Network (PSPNET2), ensemble deep learning, and different bands such as ash color scheme using Bands 11, 14, 15, and 08, this study investigates their collective impact in improving contrail identification. Results found that the selected deep learning model significantly affected the detection process, but incorporating band 08 into the input channel did not significantly improve model performance. The most effective model was the FCN equipped with Band 08, with the lowest average loss during training (0.032591) and validation (0.013321). This research is expected to improve contrail detection in satellite imagery by using deep learning models and band selection to assist policymakers and researchers in developing strategies to reduce aviation climate impacts.