Abstract: Contrails are estimated to be the largest contributor to aviation’s net climate impacts. Avoiding the production of contrails by rerouting aircraft around contrail forming regions could reduce this impact without needing the development of new technologies, and at the cost of a marginal fuel-burn increase.

Model- or observation-based contrail avoidance strategies require the prediction of contrail forming regions to be accurate at the scale of individual flights. Robust contrail detections methods are necessary both for observation-based forecasts and for improving model-based approaches. However, current approaches for contrail detection are often inconsistent from minute to minute, resulting in inconsistent forecasts of contrail formation which are difficult for aircraft to work around.

We resolve this issue by applying an ensemble Kalman filtering (EnKF) approach with an existing deep-learning contrail detection framework which identifies contrail pixels on geostationary satellite imagery. The EnKF increases the robustness of the detections by representing the temporal correlation between consecutive detections, thereby enabling consistent identification of contrail forming regions.

We evaluate the performance of the EnKF against a hand-labeled dataset of over 70 contrails tracked over a two-hour period. On average, we find that after filtering, we increase both the number of contrail pixels recovered on an image, and the number of pixels correctly predicted as contrail pixels. By adding temporal correlations, we successfully increase the duration over which a given contrail is detected consistently. The improved robustness of the contrail detections enables more consistent observation-based contrail forecasting, as well as the tracking of individual contrails. These tracks are used to derive the evolution of contrail properties such as lifetime at the individual scale. This will allow for direct comparisons between contrail models and observational data.