Presentation Information
[O12-P86]Development of an Automated Sprite Observation System Using Machine Learning
*Kitamura Ryuki1, *Oishi Yuki1, *Nemoto Miyu1, *Suzuki Hideto1, *Hirakawa Mizusa1 (1. Tokyo Metropolitan Tachikawa Senior High Schoo)
Keywords:
Sprite,Machine Learning,Lightning
[Introduction]
Sprites are highly transient atmospheric phenomena caused by tropospheric lightning discharges influencing the upper atmosphere (Fig. 1). Elucidating their mechanisms is crucial, but their short duration and low occurrence frequency make detailed analysis challenging. After capturing 28 sprites using our school's all-sky meteor camera (Fig. 2), we found its low temporal resolution and monochrome sensor inadequate. Therefore, we developed a dedicated observation system using machine learning to automatically predict occurrence regions and classify images. Furthermore, based on previous studies, we hypothesized that sprites are generated by electron migration from thunderclouds to the ionosphere, and we investigated this correlation using ionograms.
[Methods]
We analyzed meteorological conditions (precipitation, lightning, pressure patterns) for 8 well-documented sprites recorded by the security camera (Fig. 3). Estimated sprite locations were plotted on a map using reports from the amateur observation network "Sonota.Co" (Fig. 4). We built an automated tracking system with a 2-axis servo motor mount (Figs. 5–9). Using real-time data from the "Blitzortung" lightning network, it targets active thunderstorms. We developed a deep learning model to extract luminous regions and automatically detect sprites (Fig. 10), implementing an algorithm to filter out false positives caused by the apparent movement of artificial lights during camera rotation. Sprite locations were estimated using camera pointing angles and their positions in images. Based on a prior research's hypothesis that the horizontal distance between a sprite and its parent lightning is proportional to their time delay, we formulated an approximation to calculate spatial relationships. Using this equation and the pointing direction of the observation system, we estimated sprite locations (Fig. 11). Using these estimations, we analyzed surrounding atmospheric conditions and utilized ionograms to investigate the relationship between sprites and the ionosphere (Fig. 12).
[Results and Discussion]
All 8 initial cases involved precipitation and lightning during winter (6 over the Sea of Japan, 2 over the Pacific). Occurrences primarily over the Sea of Japan align with prior research. Pacific occurrences were noted, but observation bias (clearer skies, varying equipment) complicates direct comparison. The custom system captured 13 sprites but showed a high miss rate. Since static image detection was highly accurate, this limitation is attributed to hardware constraints (camera sensitivity, frame rate, environmental resistance) rather than software (Fig. 13). We analyzed 11 out of 43 observed sprites. For 5 cases, a single parent lightning strike was successfully identified; for 6, multiple candidates remained. Our approximation formula assumed a simple time-distance proportionality, ignoring electric field structures and Earth's curvature. Time delay calculations also suffered from 1/100-second limits in equipment resolution. Consequently, our estimates showed up to a 15% relative error compared to Sonota.Co's multipoint observations. Analyzing ionograms for 14 events (11 from our observation + 3 from Sonota.Co), 11 showed relatively active ionospheric conditions near 100 km altitude during occurrences. However, the limited data precludes definitive statistical conclusions.
Sprites are highly transient atmospheric phenomena caused by tropospheric lightning discharges influencing the upper atmosphere (Fig. 1). Elucidating their mechanisms is crucial, but their short duration and low occurrence frequency make detailed analysis challenging. After capturing 28 sprites using our school's all-sky meteor camera (Fig. 2), we found its low temporal resolution and monochrome sensor inadequate. Therefore, we developed a dedicated observation system using machine learning to automatically predict occurrence regions and classify images. Furthermore, based on previous studies, we hypothesized that sprites are generated by electron migration from thunderclouds to the ionosphere, and we investigated this correlation using ionograms.
[Methods]
We analyzed meteorological conditions (precipitation, lightning, pressure patterns) for 8 well-documented sprites recorded by the security camera (Fig. 3). Estimated sprite locations were plotted on a map using reports from the amateur observation network "Sonota.Co" (Fig. 4). We built an automated tracking system with a 2-axis servo motor mount (Figs. 5–9). Using real-time data from the "Blitzortung" lightning network, it targets active thunderstorms. We developed a deep learning model to extract luminous regions and automatically detect sprites (Fig. 10), implementing an algorithm to filter out false positives caused by the apparent movement of artificial lights during camera rotation. Sprite locations were estimated using camera pointing angles and their positions in images. Based on a prior research's hypothesis that the horizontal distance between a sprite and its parent lightning is proportional to their time delay, we formulated an approximation to calculate spatial relationships. Using this equation and the pointing direction of the observation system, we estimated sprite locations (Fig. 11). Using these estimations, we analyzed surrounding atmospheric conditions and utilized ionograms to investigate the relationship between sprites and the ionosphere (Fig. 12).
[Results and Discussion]
All 8 initial cases involved precipitation and lightning during winter (6 over the Sea of Japan, 2 over the Pacific). Occurrences primarily over the Sea of Japan align with prior research. Pacific occurrences were noted, but observation bias (clearer skies, varying equipment) complicates direct comparison. The custom system captured 13 sprites but showed a high miss rate. Since static image detection was highly accurate, this limitation is attributed to hardware constraints (camera sensitivity, frame rate, environmental resistance) rather than software (Fig. 13). We analyzed 11 out of 43 observed sprites. For 5 cases, a single parent lightning strike was successfully identified; for 6, multiple candidates remained. Our approximation formula assumed a simple time-distance proportionality, ignoring electric field structures and Earth's curvature. Time delay calculations also suffered from 1/100-second limits in equipment resolution. Consequently, our estimates showed up to a 15% relative error compared to Sonota.Co's multipoint observations. Analyzing ionograms for 14 events (11 from our observation + 3 from Sonota.Co), 11 showed relatively active ionospheric conditions near 100 km altitude during occurrences. However, the limited data precludes definitive statistical conclusions.