Classifying Cloud Top Phase with GOES-16 Longwave Bands, ICICLE Data, and Machine Learning

The Terminal Area Icing Weather Information for NexGen (TAIWIN) project focuses on delivering icing weather information to aircraft operators in and around the terminal area to improve the safety and efficiency of aircraft operations. Geostationary satellite data from the GOES-16 Advanced Baseline Imager (ABI) is an important component of the TAIWIN capability, particularly noteworthy for its role in determining cloud top parameters such as cloud top phase. The present cloud top phase determination capability uses a three-branch processing tree delimited by solar zenith angle (SZA) to account for the sensitivity of ABI shortwave bands to reflected solar radiation. The resulting cloud top phase answer is well-tuned to SZA but also exhibits sharp transition lines between condition branches which manifest as non-physical boundaries in downstream parameters. This study explores machine learning avenues to classify cloud top phase in a single branch using only brightness temperatures and brightness temperature differences across the nine longwave bands of the ABI. Cloud top phase training and validation data were collected during the FAA In-Cloud Icing and Large Drop Experiment (ICICLE) by instruments aboard the NRC Convair 580. Skill scores for classification of liquid, ice, and mixed phase cloud tops via a multi-layer perception are shown to exceed 80% in many cases and are examined in detail. The limited scope of sampled cloud types and other dataset limitations are discussed, as well as possible solutions to these limitations.

This research is in response to requirements and funding by the FAA. The views expressed are those of the authors and do not necessarily represent the official policy of position of the FAA.