Springtime in the Northern Hemisphere marks the onset of tornado season, and while the dust and debris-filled spiral of a tornado may seem an unmistakable sight, these violent weather phenomena often evade detection until it’s too late. Recognizing the need for better ways of predicting these occurrences, researchers at MIT Lincoln Laboratory have compiled a dataset containing radar images from thousands of tornadoes which have struck the U.S. over the past decade in the hope of enabling breakthroughs in the detection and forecasting of tornadoes through machine learning algorithms.
Released open source and named TorNet, the dataset contains more than 200,000 radar images, approximately 13,587 of these featuring tornadoes. The rest of the images are non-tornadic and include images from severe storms randomly selected, and from false-alarm storms, i.e., storms which appeared severe enough to issue a tornado warning but ultimately did not materialize into anything.
Researchers Mark Veillette and James Kurdzo have painstakingly curated the dataset ensuring a balance of both tornadic and non-tornadic samples. “If the dataset were too easy, say by comparing tornadoes to snowstorms, an algorithm trained on the data would likely over-classify storms as tornadic,” Kurdzo said.
Simultaneously, predictive models have been developed and trained based on the data in TorNet. As Veillette explains, a ground-breaking dataset can drive significant progress as it standardizes the data used by researchers, fostering commonality and ease of comparison in the meteorological and data science disciplines. The notable feature of the models based on this dataset is the use of artificial intelligence (AI) to identify key features and patterns across the radar imagery essential for detecting the onset of tornadoes. The performance of the AI-driven deep learning model was comparable to, if not better than, earlier models used for tornado detection.
While it is still early days, the researchers have a vision for the application of these models. They believe that AI could potentially help unravel the mysteries of tornadic formation, supplementing human forecasters, and providing timely visual warnings to areas predicted to experience tornadic activities.
This unprecedented move towards the utilization of AI in the detection of these destructive weather phenomena represents a significant stride in ensuring public safety. “We may never get more than a 10- to 15-minute tornado warning using these tools. But if we could lower the false-alarm rate, we could start to make headway with public perception,” Kurdzo comments on the potential. He pins his hopes on the global community of researchers who, inspired by TorNet, will develop their own algorithms that will gradually transition into operational use. With this advancement, the path to better forecasting and detection of tornadoes seems bright, fostering crucial trust in weather forecasting technologies.
