Skip to main content

Machine Learning in Analyzing Atmospheric Gravity Waves

Anastasia
Brown
First Author's Affiliation
Utah State University (USU) Physics Department Center for Atmospheric and Space Sciencess (CASS)
Abstract text:

The ANtarctic Gravity Wave Instrument Network (ANGWIN) is an international collaboration aimed at investigating the upper atmosphere dynamics over a continent-size region, using a network of all-sky imagers. In an effort to streamline the identification of “clean” windows in the extensive database of all sky-imager data obtained since 2012, we have developed a machine learning algorithm that sorts “clean” (marked as 0) images from “obscured” (marked as 1) images. Already, we have successfully created a LightGBM (Light Gradient-Boosting Machine) model that accurately sorts through images taken at the Davis and McMurdo research stations in Antarctica. The LightGBM model for the Davis station reported a 98.5% accuracy (or certainty), and the quality of the images found were high enough to use in airglow analysis. The LightGBM model for the Halley station reports a 99.2% accuracy, but when examining Halley data from 2012, in the months of August and July alone, 43 out of 69 “clean” windows were wrongly mislabeled and were “obscured”. To improve the Halley LightGBM model, the misidentified windows were combed through to identify weak points in the model. Once identified, prime examples of the mislabeled data were added to Halley’s Machine Learning training and testing set labeled as “obscured” to improve the model. A second Halley model was made and reports a 98.7% accuracy and when examining one month’s worth of data correctly identified 10 out of 13 clean windows. Further testing of the second Halley model needs to be done, but already its ability to correctly identify clean windows has greatly improved.

Student in poster competition
Poster category
MLTG - Mesosphere and Lower Thermosphere Gravity Waves