Researchers at MIT Lincoln Laboratory have developed a new open-source dataset, named TorNet, to detect and predict tornadoes. By using artificial intelligence (AI) models trained on TorNet, researchers hope to improve tornado forecasts and warning accuracy, potentially saving lives and minimizing damage.
Tornadoes are challenging to predict, and this represents a high false alarm rate – over 70%, leading to a potential “boy-who-cried-wolf” situation. The TorNet dataset attempts to mitigate this problem leveraging over a decade of radar data from thousands of US tornadoes. The dataset is substantial, with more than 200,000 radar images, 13,597 of which depict tornadoes, while the rest are classified as non-tornadic, from randomly selected severe storms or false-alarm storms.
The researchers used the TorNet dataset to build baseline AI models and found an excellent potential in deep learning, a machine learning subfield excellent for processing visual data. The outcomes are promising, with the deep learning model performing as well as or better than existing tornado-detecting algorithms. The model accurately classified 50 percent of weaker EF-1 tornadoes and over 85 percent of tornadoes rated EF-2 or higher.
As the TorNet dataset is open-source, it encourages other researchers worldwide to use it to build and improve upon the proposed deep learning models. In addition to predicting tornadoes, these models might help unearth insights into why tornadoes form in the first place.
Although deploying operational algorithm-based detection is a long road due to safety considerations, the researchers believe that public benchmark datasets like TorNet represent an essential first step. Kurdzo, one of the co-principal investigators of the project, even believes that while these tools may not be able to extend tornado warnings beyond 10-15 minutes, they could significantly reduce false alarms and improve trust in such warnings, encouraging people to take appropriate lifesaving measures.
The project was funded by MIT Lincoln Laboratory’s Climate Change Initiative, which aims to leverage the laboratory’s diverse technical strengths to address climate problems threatening global security and human health. This data science-meets-climate science approach has already shown promise and offers exciting potential for the future of meteorology.