Disentangling Representations of Ice Crystals with In Situ Imagery and Unsupervised Learning

Ice clouds are difficult to model due to large (spatial and temporal) scale disparities and an incomplete understanding of microphysics. A key ice cloud microphysical property is particle habit (i.e., shape). At the single-particle scale, habit influences processes such as growth, collection, radiation interactions, and settling velocity. At the bulk-cloud scale, habit influences the radiative transfer, spatial distribution, smaller scale dynamics and ultimately the evolution of ice clouds. One way we can constrain the representation of ice crystal habit in models 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 this work, we leverage recent advances in representation learning to unravel the underlying diversity of ice crystal habit from in-situ optical imagery. Bengio et al. (2014) define representation learning as “...learning representations of the data that make it easier to extract useful information when building classifiers or other predictors.” With this foundational definition in mind, we narrow the scope of our work to two main goals. (1) First, we use variational autoencoders and the k-means algorithm to cluster CPI images in an unsupervised manner. We find that clustering CPI images using their latent representations rather than raw images drastically improves the results of downstream clustering. (2) Second, we attempt to disentangle and interpret the underlying factors of variation in the latent representation of ice crystals. To do this, we experiment with methods of “disentangled representation learning” (Wang et al. 2022) developed largely by the computer vision and image processing community. Incorporating representation learning into the large-scale analysis of in-situ ice crystal imagery will help establish an observationally-constrained understanding of ice habit in real-world clouds, and ultimately improve ice process representations in atmospheric numerical models.