Classification of Cloud Particle Imagery and Thermodynamics (COCPIT): A New Databasing Tool for the Characterization of Cloud Particle Images Captured During DOE Field Campaigns
Recent computational achievements have provided the unique opportunity to perform comprehensive data-driven analysis on a multiplex of big-data applications within the atmospheric sciences and beyond. Of late, widespread success surrounds machine learning (ML) and other big-data statistical techniques as powerful tools for efficiently operating on large datasets. Specific to image classification, ML algorithms can select important features and pattern repetitions that could otherwise go unnoticed.
The suite of DOE field campaigns and associated aircraft instrumentation present a unique opportunity to exploit these emerging methods to glean scientific insight previously not possible. In particular, this work proposes to cull DOE datasets of Cloud Particle Imager (CPI) probe images from Intensive Operation Periods (IOPs, on the order of millions of datapoints) to develop a robust methodology for particle image categorization. CPI imagery is plentiful and broad but has yet to be assembled into a coherent interface and is under-utilized in a bulk and statistical sense. The high-resolution imagery and environmental properties associated with each ice particle habit (shape) from a suite of in situ aircraft probes make for an invaluable testbed for cloud modeling, radiative forcing, and climatic thermodynamic feedbacks across all latitudes. A Convolutional Neural Network (CNN) will be trained on these CPI images to identify a given particle type (e.g., aggregates, rimed aggregates, columns, rimed columns, spheres or liquefied drops, probe shattered fragments, etc.). Concurrently, (1) a Particle Image Characterization Tool (PICT) tool will be used to extract dimensional and other particle properties from CPI images, and (2) the environment in which these particles were captured gleaned from accompanying IOP aircraft probes will be defined. Further, understanding that associated environmental information does not define particle growth history, HYSPLIT (Hybrid Single Particle Lagrangian Integrated Trajectory Model) methodologies will be used to provide information on the temporal history of the thermodynamics associated with the observed crystal habits. The results of PICT, the ML model, direct environmental mapping, and HYSPLIT trajectories will be inserted into a database. Scripts will be developed to enable users the ability to insert, access, and analyze from the database, with filtering capabilities. This suite of scripts and the database comprise COCPIT: A tool for the Classification of Cloud Particle Imagery and Thermodynamics, and will be made publicly available to users via the DOE ARM archive and the DOE GitHub repository.
With the creation of a habit and characteristic classifier that achieves high-reaching accuracies on unseen data and the addition of collocated environmental properties, robust statistics on a wide-ranging spatial and temporal extent will be achievable. Past studies on ice crystal habit required a time consuming pre-processing phase; a user-friendly data product with greater automation and filtering capabilities would direct scientist time sensitivities toward analysis and conclusions of ice microphysical and thermodynamic processes instead of mulling over data practices. While other past classification efforts have proven successful, a convenient interface that captures boundless data points with collocated in situ environmental properties is not readily available. Such a database would limit ‘cherry-picking’ or biased sampling for a given regime, improve mass and area-dimensional relationships, confine aspect ratios for dual-polarization radar verification, and ultimately improve parameterizations of remote retrievals and numerical simulations through a better understanding of growth regimes.
Przybylo V, K Sulia, C Schmitt, and Z Lebo. 2022. "Classification of Cloud Particle Imagery from Aircraft Platforms Using Convolutional Neural Networks." Journal of Atmospheric and Oceanic Technology, 39(4), 10.1175/JTECH-D-21-0094.1.