Towards Global Contrail Observation: From Single-Instrument to Global Geostationary Data

One of aviation's largest non-CO2 climate impacts originates from persistent contrails. Since only a small minority of flights cause persistent contrails, preventing contrail formation has potential to be cost-effective in comparison to other measures and therefore could be part of climate change mitigation strategies in the transport sector. However, determining which flights cause persistent contrails to form is uncertain, as is the overall extent of the impact of contrails on climate change. More research is needed in both areas to fully understand and mitigate these uncertainties.

Observations of persistent contrails through geostationary satellite-based imagers can be used to develop non-CO2 climate impact inventories. Recent work has focussed on training machine learning models to detect contrails at large scale. To date, these models have been trained and evaluated on observations from the same satellite instrument. But in order to get global contrail coverage, one must consider multiple instruments mounted on different satellites (e.g., GOES-ABI for the Americas, Meteosat-FCI for Europe and Africa, and Himawari for Asia). 

In this work, we analyze whether a deep learning model trained on one satellite instrument can be applied to data from others. Validating this approach is important, as it could eliminate the need to create large labeled datasets for every new instrument which is a time-consuming and expensive process. We also explore if training models on a combined dataset of multiple satellite instruments can lead to overall quality improvement in contrail detection.