Satellite-based Analysis and Forecast Evaluation of Aviation Contrails

Abstract

Theclimate impact of aviation is currently estimated to account for 3.5% of anthropogenic climate forcing when quantified in terms of effective radiative forcing. Non-CO2 impacts are thought to represent over half of this aviation-induced forcing, with contrails and contrail cirrus being the largest contributor. Contrails (short for condensation trails) are the lineshaped clouds that form as a result of the warmer, moister engine exhaust mixing with the colder ambient air. When the air is ice supersaturated, a formed contrail can persist for multiple hours and become indistinguishable from natural cirrus. Contrails interact with Earth’s energy balance by reflecting incoming shortwave radiation (a cooling effect) and absorbing outgoing longwave radiation (a warming effect), with the net effect estimated as warming. The ice supersaturated regions that allow for contrails to persist are known to be horizontally wide but vertically thin, which has motivated the idea of contrail avoidance. Small changes in altitude, with minimal associated fuel burn penalty, may be sufficient for avoiding areas where contrails can persist. This concept has been extensively studied by means of simulations, which all indicate that the avoided warming by contrails far outweighs the costs and climate-impacts of the additional fuel burn. However, all of these studies assume that the location of ice supersaturated regions is known perfectly, and they therefore ignore the impact of forecasting accuracy on the effectiveness of contrail avoidance. With recent studies indicating that numerical weather prediction models have limited ability to predict ice supersaturation, there is a pressing need for further evaluation of the capability of models to predict contrail persistence as well as their improvement. Despite efforts made in the past and ongoing, there is no objectively evaluated implementation of contrail avoidance at the time of writing. This doctoral thesis develops an approach to locate contrails in satellite imagery and utilize these observations to assess existing methods for contrail prediction, as well as provide improved short-term forecasts. Contrails are found in geostationary GOES-16 ABI imagery 2by use of a convolutional neural network. Once found, their altitude is estimated using this same imagery by another deep learning algorithm that was trained on contrails collocated in data from the satellite-based CALIOP LIDAR. A dataset for the purpose of contrail forecast evaluation is developed by use of an algorithm that finds the location of aircraft exhaust plumes and contrails within CALIOP data. The resulting data provides observations of persistent contrail formation at the flight-by-flight level. Existing prediction methods that use numerical weather prediction (NWP) data from the ERA5 and HRRR models are compared to a nowcasting method that utilizes the aforementioned contrail detections and altitude estimates. This comparison is performed using traditional forecast evaluation methods, as well as a novel framework developed specifically for assessing the achievable benefits of contrail avoidance. The developed contrail detection algorithm is applied to more than 100,000 GOES-16 ABI satellite images to study contrail coverage over the United States for the years 20182020, and how it relates to changes in flight traffic density. This effort represents the first long-term (> 2 years) study of contrail coverage using satellite imagery. It is also the first extensive application of a contrail detection algorithm to geostationary satellite data, such that estimates of contrail coverage are available at a temporal resolution of 10-15 minutes. This is in contrast to previous studies that detected contrails in low-Earth orbit (LEO) satellite imagery where a particular region is only visible twice a day. As a result, it is possible to directly study the diurnal patterns in contrail coverage and how these relate to f light activity. Furthermore, the year 2020 included the unprecedented reductions in flight activity as a consequence of the COVID-19 pandemic. This provides a unique opportunity to study how contrail coverage changed with large reductions in flight traffic. The contrails detected in the GOES-16 ABI imagery are then collocated with data from the CALIOP LIDAR aboard the CALIPSO satellite. The resulting dataset provides crosssectional and altitude information on more than 3000 contrails. This dataset is used to create and assess multiple algorithms for the estimation of contrail altitude by use of GOES-16 ABI imagery alone. The best performing algorithm is a convolutional neural network that is first trained on CALIOP cirrus data and then fine-tuned on the contrail data. This algorithm has a root mean square error of 570 meters and a coefficient of determination (R2) of 0.76. These results suggest that image-level context is useful for estimating contrail altitude, and that training on contrail data is better than training on cirrus data alone as was done in previous work. The altitude estimation model also provides an estimate of its uncertainty: the 95% confidence intervals (CIs) derived from the model’s output are shown to contain approximately 95% of the test data points and are narrower (2.0 km on average) than such intervals derived from flight data near the observed contrails (4.4 km on average). Finally, a method to match contrails found in CALIOP data to the flights that formed them is used to create a dataset for the purpose of evaluating persistent contrail forecasts at the flight-by-flight level. This dataset is used to evaluate the performance of predictions of ice supersaturation from ERA5 and HRRR as well as a nowcasting approach that relies on the contrail detections and altitude estimates developed here. The best performing nowcast approach is able to achieve higher hit rates at low false alarm rates (1 to 15%) than predictions using the ERA5 and HRRR relative humidity w.r.t ice. However, the nowcasting approaches have upper limits on their hit rates (< 68%) as their predictions are constrained to areas where contrail are observed. A framework to assess the benefits that can be achieved by 3using a particular forecasting system for contrail avoidance is developed as well. It shows that existing forecast performance metrics cannot directly be used to evaluate the suitability of a particular prediction method for contrail avoidance, as the forecast needs to be ‘correct twice in a row’: at the original and the deviated aircraft flight level. The results obtained using this new framework indicate that, depending on the relative benefits of avoiding a contrail and costs of additional fuel burn, the nowcasting approach may lead to a more effective contrail avoidance implementation than possible with current NWP data.