Every spring, tornado season returns to the Northern Hemisphere. While the twisted funnel of a tornado may seem like an easily recognizable sight, it remains difficult for radar — the primary tool of meteorologists — to detect as they form. Predicting tornadoes also remains challenging, due to an unclear understanding of why they form. Over the past ten years, there are around 1,200 tornadoes that occur annually across the United States, causing significant economic damage and an average of 71 lives lost each year.
To shed light on this natural disaster mystery, researchers from MIT Lincoln Laboratory curated a new dataset called TorNet. The open-source dataset contains radar returns from thousands of tornadoes over the past decade, and the storms that birthed them as well as storms with identical conditions that did not generate tornadoes. The goal of TorNet is to provide machine learning algorithms with a benchmark dataset to improve the detection and prediction of tornadoes.
The released dataset contains over 200,000 radar images, with 13,587 featuring tornadoes. The rest of the images are from non-tornadic storms, either randomly chosen severe storms or those that resulted in false alarms. These images have been taken from two different radar sweep angles, featuring diverse radar data products such as reflectivity for precipitation intensity and radial velocity for wind direction.
The data’s curation presented challenges as finding tornadoes within weather radar data was difficult due to their extreme rarity. The researchers had to strike a balance between tornado samples and difficult non-tornado samples. If the dataset comprised only the easy non-tornado samples, it could over-classify storms as tornadic, weakening the resulting algorithm’s relevance.
The researchers used this dataset to develop baseline artificial intelligence (AI) models and were keen to leverage deep learning, a form of machine learning highly effective in extracting features from images across datasets. Their deep learning model’s performance was promising, detecting 50% of weaker EF-1 tornadoes and over 85% of EF-2 tornadoes and above. For comparison, the researchers also evaluated two other types of machine-learning models and one traditional model.
While the primary utility of this dataset is to improve the reliability of tornado predictions and warnings, it could also resolve the science behind tornadoes’ formation. Gonzo hopes that “explainable AI” – AI models that can explain their decision-making in a human-understandable manner, might help identify precursors to a tornado and train forecasters to recognize signs at an earlier stage.
The researchers hope that TorNet and the associated models will incentivize global researchers to develop their algorithms, leading to better detection and prediction of tornadoes. They are especially optimistic about deep learning’s potential in real-time monitoring of radar returns as data refresh rates increase. However, transitioning these algorithms into operations will take time, primarily amidst the forecaster community’s skepticism of machine learning. The researchers believe that public benchmark datasets like TorNet can help establish trust and transparency, and eventually, help people take the necessary action to save their lives during a tornado.
