Incubit Improves Tornado Detection AI in Mountainous Area as Contractor for Meteorological Research Institute
Incubit, as a contractor for MRI, has improved its tornado detection AI system by reducing over-detection in mountainous areas, allowing for a more robust system that is not limited by radar location.
TOKYO, June 30, 2021 /PRNewswire-PRWeb/ -- Incubit, Inc. (hereafter: Incubit), which undertakes practical implementation of image recognition by utilizing deep learning technologies, is working to train an AI model to automatically detect tornadoes. By adding topographic AI data alongside Doppler radar data, the company has been able to reduce over-detection of tornadoes in mountainous areas. This development was made under the 2020 research and development project, "AI Based Tornado and Heavy Localized Rain Prediction and Information Distribution System", which the Meteorological Research Institute (hereafter: MRI), which works directly under the Japan Meteorological Agency as an official research institution, was in charge of. This outcome brings hope for a robust tornado detection AI model, the use of which is not limited by its location. This outcome provides prospects for further research plans in 2021, where a variety of different data sets will be used for training to increase the accuracy of an already precise model.
There are, on average, more than 20 tornadoes occurring in Japan each year, blowing down houses and derailing trains. MRI has been working on utilizing deep learning technology to automatically detect tornado incidents since 2018.
The objective of this particular research, that MRI and Incubit conducted in 2020 was to reduce over-detection in mountainous areas by using altitude and topographic data as part of the training data, so that geographical traits could be incorporated by the tornado detection AI into its decision-making process.
Doppler radar uses the reflection wave of precipitation particles in the air to find the direction and speed of the moving wind. To the trained eye, tornadoes have a very particular vortex pattern detectable using Doppler radar. It was later found that in mountainous areas, the geographical features interfered with the reflection of the Doppler radar, not only making the pattern different between mountainous areas and the plains, but creating patterns that mimic those of tornado vortices, which confused the AI model to the point that it over-detected tornadoes in mountainous areas.
After the research conducted in 2020, Incubit and MRI were able to conclude that it is possible to reduce over-detection by using topographical data in training, keeping the average detection rate of real tornadoes high while reducing over-detection by 50%. This allows for the possibility of Doppler radars in mountainous locations to be used for automatic tornado detection.
Incubit and MRI plan to further expand on this success by adding various different datasets in 2021 for the model to train on, including timeline data, heavy rainfall areas, and wind data. The final goal is to utilize the tornado detection system in transportation systems such as trains, cars and even drones, for a safer and more autonomous transportation system.
*This work was supported by Cabinet Office, Government of Japan, Public/Private R&D Investment Strategic Expansion Program (PRISM) in 2020, and will be continued in 2021
About Incubit
Incubit undertakes the challenge of solving industry specific problems alongside the leading companies of Japan adopting the latest technology. We've worked in the fields of medicine, bio tech, space, geology and manufacturing with notable achievements, by virtue of our earnest approach towards solution deployment.
Our current focus and strength is image recognition using deep learning technology, nevertheless, our enthusiasm always lies in finding the best possible solution providable.
Media Contact
Julia Ryan, Incubit Inc., +81 364502377, [email protected]
SOURCE Incubit Inc.
Share this article