Improving the temporal consistency of satellite-based contrail detections using ensemble Kalman filtering

Abstract

Condensation trails or contrails are line-shaped ice clouds which can form behind aircraft, and current estimates indicate that they account for the majority of aviation’s climate impacts. While contrail models exist to estimate these effects, a lack of experimental or observational data makes them difficult to validate.

This thesis develops a method for retrieving large scale temporally consistent observations of contrails using satellite imagery. Having a consistent history of detections of an individual contrail is necessary to accurately derive observational constraints on contrail properties such as lifetime. Inconsistencies not only reduce the quality of such a dataset, but risk introducing biases in the computed properties.

We use an existing deep-learning based contrail detector which as of now presents temporal inconsistencies that make tracking challenging. We address this issue by post-processing the model’s outputs with an ensemble Kalman filter. We create a hand-labeled dataset of 73 contrails tracked over a 2-hour time series which we use to quantify performance. We find that by adding temporal correlations, we are able to recover 53.25% of contrail pixels on an image, and that 53.25% of the pixels predicted as contrail by the detection framework are indeed contrail pixels. For individual contrail tracks, we find that after filtering, we increase the average duration of consecutive consistent contrail detections from 9.4 minutes to 25.7 minutes. On average, the duration of these consistent contrail detections after filtering represent 43.7% of a contrail’s total lifetime compared to only 15.5% for the baseline. We also find that the high frequency Fourier components of the signal, which are responsible for flickering and noise, are reduced by 50% in magnitude.