Evaluation and Comparison of Machine Learning Techniques for Improving Low-Level Polar Cloud Retrievals

Operational space-based retrievals of polar cloud properties can be problematic due to a low signal-to-noise ratio (SNR) in the visible and infrared channels typical of a wide field-of-view sensor like VIIRS. However, recent work has shown that automatic pattern recognition using machine learning techniques may help to increase SNR for detecting clouds (Haynes et al., 2022; Shao et al., 2019; White et al., 2021; Yu and Lary, 2021), distinguishing cloud types (Marais et al., 2020), and retrieving cloud properties (Min et al., 2020; Wang et al., 2020). Here we focus on improving the retrieval of low-level Arctic clouds and their properties by training VIIRS and other polar-orbiting imagery datasets with ground-based observations from the North Slope of Alaska and MOSAiC Expedition data from the Central Arctic. We evaluate and compare the performances of several machine learning model architectures, including various neural networks and a random forest classifier. Specifically, we compare the ability of these models to accurately characterize not only the presence of these low-level clouds, but also their top and base heights, phases, and the presence of precipitation.