An Evaluation of a Convolutional Neural Network for Classifying Images from In-situ Cloud Probes

A vast amount of ice crystal imagery exists from the Particle Habit Imaging and Polar Scattering (PHIPS) probe that has been deployed on a variety of field projects to obtain in-situ cloud microphysical observations. During a cirrus anvil experiment near Cape Canaveral, Florida in 2019 (CapeEx19), a vast amount of ice crystal imagery was obtained where a large percentage of particles were classified as chain aggregates. Chain aggregates are defined with at least one of the following characteristics: three or more discernible particles oriented in a quasi-linear fashion, particles joined together by small joints, and links of particles that are unusually elongated. Classifying crystals by habit provides information about their origin since different habits form under different temperature and humidity conditions. Traditional classification methods where images are manually reviewed show reasonable performance; however, it requires a large amount of a scientist's time. Given the sizable data set gathered during recent field projects such as CapeEx19 and NASA's Investigation of Microphysics and Precipitation for Atlantic Coast-Threatening Snowstorms (IMPACTS) field projects, there is a need for an automated classification approach. Convolutional Neural Networks (CNN) have the potential to reduce the time required by using a training data set produced by manual classification to create a complex nonlinear model directly from the images. The performance of a CNN model to classify particles by habit, including chain aggregates, is evaluated. The approach uses a data-driven model to classify the PHIPS images gathered during the CapeEx19 and IMPACTS field projects with an attributed confidence. To train the model, a manually labeled data set of images from the CapeEx19 and IMPACTS fields is used. An evaluation data set is used to test the developed network and evaluate its performance.