Japan Geoscience Union Meeting 2018

Presentation information

[JJ] Oral

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

[M-GI26] Earth and planetary informatics with huge data management

Mon. May 21, 2018 10:45 AM - 12:15 PM 201A (2F International Conference Hall, Makuhari Messe)

convener:Ken T. Murata(National Institute of Information and Communications Technology), Takeshi Horinouchi(Faculty of Environmental Earth Science, Hokkaido University), Rie Honda(高知大学自然科学系理工学部門, 共同), Susumu Nonogaki(Geological Survey of Japan, National Institute of Advanced Industrial Science and Technology), Chairperson:Nonogaki Susumu, Murata Ken(National Institute of Information and Communications Technology)

12:00 PM - 12:15 PM

[MGI26-12] Detection of information related to disasters from monitoring cameras – detection of tornadoes or funnel clouds -

Yuki Fujii1, Keita Higaki1, Shinya Nakayama1, *Rie Honda2, Koji Sassa2 (1.Department of Information Science, Kochi University, 2.Department of Science, System of Natural Science, Kochi University)

Keywords:tornado, funnel cloud, deep learning, disaster, footage, data mining

The disaster-related images or footages have been recorded by the observer who happened to be the points of disaster. The development of the information technology has enabled the real-time monitoring of images of the disaster related targets such as volcanoes or meteorological phenomenon. On the other hand, improvement of computer vision and machine learning method such as deep learning has enabled the implementation of visual recognition ability comparable to human visual recognition to the computer.
Tornadoes and funnel clouds occur at the high frequencies around the Tosa gulf in Kochi. We have expanded twelve monitoring cameras around Tosa gulf to detect the tornadoes or funnel clouds to promote studies of the mechanism of the tornadoes and for development of just before occurrence of tornadoes predicting system.
In this study, detection of funnel clouds from the monitoring camera images using convolutional neural networks (CNN) is examined. Two type of CNN for object detection and classification are examined. Furthermore, apparent velocity distribution was obtained by using optical flow to recognize the sign of tornado occurence by cloud movement. These methods are applied to the real footages on September 11, 2017, in which multiple funnel clouds are observed. The effectiveness of the proposed method is confirmed through the experiments using real images.