The arrival of spring in the Northern Hemisphere brings with it the commencement of tornado season. Meteorologists use radar to track these dangerous natural phenomena, but understanding exactly when a tornado has formed or why can be a challenge. However, a new dataset may provide some answers.
Known as TorNet, this dataset compiled by researchers from MIT Lincoln Laboratory contains radar returns from thousands of tornadoes that have hit the U.S. over the past decade. Armed with this data, the researchers hope to help understand and predict tornadoes better by advancing machine learning algorithms, which can both detect and predict tornadoes.
Alongside the dataset, the team has also released models built upon it, revealing potential for machine learning to detect tornadoes accurately. This could potentially help weather forecasters provide more precise warnings, possibly saving lives in the process.
Tornadoes are notoriously difficult to predict because scientists still lack a comprehensive understanding of why they form. Even with the primary tool of weather radar to monitor storm conditions, tornadoes still often go undetected. This limited view subsequently leads to forecasters issuing tornado warnings, of which over 70% end up being false alarms. This initiates a “boy-who-cried-wolf” scenario with the general populace.
To improve tornado detection and prediction, researchers have shifted to machine learning in recent years. The newly released TorNet dataset, containing more than 200,000 radar images with 13,587 portraying tornadoes, fills a crucial gap in this area. The rest of the images are non-tornadic, captured during storms that either led to a false alarm or randomly occurred severe storms.
The introduction of deep learning, a type of machine learning that excels at processing visual data, also played an instrumental role. The deep learning model showed promising results, correctly classifying 50% of weaker EF-1 tornadoes and over 85% of tornadoes rated EF-2 or higher. It has since been released open-source, available for the broader research community to improve upon it.
Longer-term, the researchers also hope the model could assist in unraveling the mystery surrounding why tornadoes form. This could uncover physical processes that occur before tornadoes. This knowledge could then be used to train future forecasters and models to detect these signs sooner.
The researchers firmly believe that technology, like the deep learning model, is not intended to replace human forecasters. But, it could guide them in complex weather situations, provide a visual warning for areas predicted to witness tornadic activity, and even process large amounts of data more rapidly than humans – aiding in real-time monitoring.
Although there’s a long way to go in the development of operational algorithms for managing safety-critical situations, the Lincoln Laboratory’s work is an important first step in building trust within the forecaster community regarding machine learning.
The team hopes researchers globally will be inspired by the dataset and will build their own algorithms leading to transitioning into operations. The ultimate goal is to improve public perception by reducing the false-alarm rates, potentially saving lives in the process.
