Japan Geoscience Union Meeting 2022

Presentation information

[J] Oral

M (Multidisciplinary and Interdisciplinary) » M-GI General Geosciences, Information Geosciences & Simulations

[M-GI35] Earth and planetary informatics with huge data management

Sun. May 22, 2022 9:00 AM - 10:30 AM 301B (International Conference Hall, Makuhari Messe)

convener:Ken T. Murata(National Institute of Information and Communications Technology), convener:Susumu Nonogaki(Geological Survey of Japan, National Institute of Advanced Industrial Science and Technology), Rie Honda(Department of Science and Technology, System of Natual Science, Kochi University), convener:Keiichiro Fukazawa(Academic Center for Computing and Media Studies, Kyoto University), Chairperson:Rie Honda(Department of Science and Technology, System of Natual Science, Kochi University), Ken T. Murata(National Institute of Information and Communications Technology)

9:15 AM - 9:30 AM

[MGI35-02] Automatic detection of hook echo by deep learning

*Koji Sassa1, Masaya Yamawaki2, Hirokaze Tanase3, Tomoya Kubo2, Rie Honda1 (1.Natural Science Cluster, Kochi University, 2.Faculty of Science and Technology, Kochi University, 3.Graduate School of Integrated Arts and Sciences, Kochi University)

Keywords:Radar Observation, Deep Learning, Tornado

Usually, we use the couplet of maximum and minimum signal of Doppler velocity to detect the parent clouds of tornadoes from Radar data. Then, the target signal for deep learning is the copulet. Doppler velocity, however, shows approaching anf/or departing from the radar, the data is biased by emvironmental wind velocity. Therefore we need to additional processis for deep learning. The present study aims to evaluate the availability of parent cloud detection by using hook echo of reflectivity data.
We used SSD as the argorism of CNN. We empl;oyed the reflectivity data of Kochi radar network as the anotation data and validation data.
It took 65 hours for deep learning but the detection time is only 0.6 sec. It shows that we can detect parent cloud at quasi-real time. For the enlarged parent cloud echo, the recall and precision are about 0.8 and 0.9, respectively.
But, it is found to need the improvement for the whole observation range of radar.