[P-EM12] Coupling Processes in the Atmosphere-Ionosphere System

[PEM12-P25] Localization and Classification of Gravity Wave Events from VIIRS Day/Night Band Satellite Imagery Using Machine Learning Techniques

*穂積 裕太^1,2、Jia Yue^1,2、Mostafa Seraj³、Wang Chenxi³、Wang Jianwu³、Purushotham Sanjay³、Miller Steven⁴ (1.The Catholic University of America、2.NASA Goddard Space Flight Center/CCMC、3.Department of Information Systems, University of Maryland, Baltimore County、4.Cooperative Institute for Research in the Atmosphere, Colorado State University)

キーワード：Gravity Waves、Airglow imaging、Machine Learning

A machine learning model was developed to detect and classify gravity wave events from Day/Night Band (DNB) images of the Visible Infrared Imaging Radiometer Suite (VIIRS) aboard the Suomi National Polar-orbiting Partnership (Suomi NPP) satellite. Spaceborne airglow imaging provides an excellent opportunity to study gravity waves, especially those with short horizontal wavelengths, from a global perspective. DNB/VIIRS has captured mesospheric airglow images since 2012, providing over twelve years of data for long-term studies. The broadband sensitivity (505–890 nm) of DNB results in significant contamination from lower atmospheric emissions, city lights, orographic features, and cloud reflections. Due to this contamination, automated systematic detection of wave events is challenging, and manual identification is impractical given the large volume of data. Therefore, a novel machine learning technique for identifying wave events is crucial to fully utilize the extensive DNB dataset for long-term research. In this study, we focus on four types of wave events with distinct appearances: concentric gravity waves, frontal waves, ripples, and other gravity wave events. These waves are associated with different physical processes, making it beneficial to distinguish among them. The YOLOv5 machine learning model, short for “You Only Look Once version 5,” was trained using manually labeled gravity wave event images with the four event labels. After training, we applied the model to twelve years of DNB data. In this presentation, we will discuss the model training process, performance evaluations, and the global distribution of the detected wave events.