A new Machine Learning-based Detection Algorithm for Potential Ice Crystal Icing Conditions from Geostationary Satellite Imagery

Kurzfassung

Ice Crystal Icing (ICI) is a condition encountered by aircraft that fly through cloudy regions with high ice crystal concentrations. These areas are often close to deep convection. Aircraft’s onboard radar can easily detect precipitating particles. Nevertheless, cloudy regions in deep convective clouds with little reflectivity can still have high ice crystal concentrations. For this reason, these areas look safe to fly but they can lead to ice accretion inside the engine combustion chamber. This eventually results in performance loss and engine damage. Because of the importance for airlines to detect these events continuously, geostationary-based algorithms for potential ICI detection were recently proposed. The input data for the development involved either research flight campaigns targeting ICI conditions or ICI events databases from aircraft manufacturers. However, some of the shortcomings of previous detection approaches were the geographical concentration of ICI events during flight campaigns and the sparse sampling in aircraft manufacturers’ ICI databases. This thesis aims to develop a novel machine learning approach for potential ICI detection exclusively using remote sensing data. High ice water content (HIWC) at typical cruise levels is a proxy for areas with potential ICI formation. The threshold for HIWC was set to 0.5 g/m³. The IWC was obtained by a combination of radar and lidar measurements (the DARDAR dataset) onboard polar-orbiting satellites CloudSat and CALIPSO. Potential predictors of HIWC conditions were taken from the geostationary satellite Metosat Second Generation (MSG). The DARDAR dataset was collocated with the MSG data. Furthermore, cloud microphysical properties were retrieved by the algorithms APICS and CiPS, while the retrieval Cb-TRAM was used for convective cell detection. Newly computed convection-related metrics were integrated into the dataset as well. The training and testing of the algorithm were performed using the summer months of 2013 and 2015. The imbalance of the HIWC and no HIWC events in the dataset was mitigated thanks to a downsampling technique. Moreover, a model shrinkage technique was applied to the original set of predictors. The predictors were clustered according to their affinity. One predictor was selected for each cluster. The final input variables set consisted of the Brightness Temperature Difference between the water vapor window at 6.2 µm wavelength and infrared channel 10.8 µm, visible channel 0.6 µm, optical thickness and four convection metrics related to the distance to the closest convective cell, area extent of the convective cells, and convection density within a radius of 100 km. The performance was obtained by repeating the test multiple times with random sample selections. The performance metrics’ median values were a probability of detection (POD) of 83 %, a false alarm rate (FAR) of 52 %, and a critical success index (CSI) of 41 %. The retrieval was further tested with a dataset of ICI events in 2016 composed by Lufthansa, which collected events in-situ during 2016. The algorithm showed detection skills for the selected events in Europe during daytime. Four out of seven events were correctly detected. Visible channel and optical thickness input variables prevented the algorithm from working in nighttime mode. The test performed with remote sensing and in-situ data proved the feasibility of this methodology for potential ICI detection. However, some shortcomings were still encountered. For example, the IWC is highly variable within a deep convective cloud but passive imagers are rather sensitive to the upper part of the cloud. Also, high solar zenith angles can hinder the algorithm’s performance because of the three-dimensional radiative effects that lead to non-reliable cloud microphysical properties estimation. Finally, it seems challenging to discriminate HIWC from moderate IWC from passive imagers. This also depends on the threshold assumption for the HIWC and the IWC sampling height.