Wednesday, 31 January 2024: 2:15 PM
338 (The Baltimore Convention Center)
In this study, we leverage variational autoencoders (VAE) and unsupervised clustering to classify images of ice cloud particles. Ice clouds impact the Earth’s energy balance and hydrologic cycle. However, ice clouds are difficult to model due to scale disparities between microphysical processes and model grid resolution, and more fundamentally, an incomplete understanding of microphysics. A key microphysical property of ice clouds is particle habit (i.e., shape). Ice habit influences particle-level processes such as growth, radiation interactions, and deposition. At the macro-scale, habit influences the radiative impacts, evolution, and lifetime of ice clouds. Given the influence of ice habit on cloud dynamics, it is crucial to represent ice crystal shapes realistically in microphysical schemes. One way we can constrain the representation of ice habit is to use in-situ measurements from airborne field campaigns. More specifically, cloud particle imagers (CPI) have been deployed in numerous airborne field campaigns, but limited analyses have been conducted using the vast dataset of millions of images. Przybylo et al. (2022) used convolutional neural networks to classify CPI images in a supervised manner, but limited work has applied unsupervised learning to CPI image classification. In our work, we propose a two-step workflow: (1) train a VAE with millions of CPI images, then (2) perform unsupervised clustering in the latent space of the VAE (e.g., using k-means, Gaussian mixture models, hierarchical clustering). We will also perform a systematic comparison between our results and results from a supervised CNN approach (Przybylo et al., 2022). Finally, we will explore the potential advantages of using unsupervised classification and discuss how habit classification can be incorporated into improved microphysical parameterizations.

