Around 1,200 tornadoes occur every year in the U.S., causing billions of dollars in damage and claiming an average of 71 lives. Predicting tornadoes is notoriously difficult due to gaps in understanding the precise conditions that cause them to form. The team from MIT’s Lincoln Laboratory hopes to address this challenge, using a new open-source dataset called TorNet. This contains radar images from thousands of tornadoes that have hit the U.S. in the past ten years.
Difficulty in predicting tornadoes arises from a reliance on weather radar, the primary tool used to monitor storm conditions. Often, tornadoes lay too low to be detected, and the rate of false alarms for tornado warnings is more than 70%, which can lead to the public disregarding warnings.
TorNet contains over 200,000 radar images, over 13,000 of which depict tornadoes. The remainder are from severe storms or false-alarm storms – those that led to a warning but didn’t produce a tornado. Each sample includes two sets of six radar images from different sweep angles, depicting data such as precipitation intensity or wind direction.
Alongside TorNet, the team is releasing models trained on the dataset to demonstrate what machine learning can bring to weather prediction. These models show promise in accurately identifying tornadoes, potentially enabling more accurate warnings and thereby saving lives. In initial testing, the model achieved impressive results, correctly classifying 50% of weaker EF-1 tornadoes and over 85% of tornadoes rated EF-2 or higher. The team believes that making these datasets and models freely available will stimulate further progress in this field.
Another potential benefit of the project could be to help unravel why tornadoes form in the first place, with the use of explainable AI. This method allows the model to provide its reasoning in making a certain prediction, shedding light on the physical processes that occur before a tornado forms. This could aid forecasters and models in identifying the early signs of a tornado. The aim is not to replace human forecasters, but to assist them in complex situations and potentially provide visual warnings of areas predicted to have tornadic activity.
Future applications of this research could involve the integration of additional data sources, such as satellite imagery or lightning maps, enhancing the accuracy of machine learning models. It could also be useful for conducting large-scale case studies on storms. The researchers hope that this dataset will inspire other researchers around the world to develop their own algorithms, with the ultimate goal of improving public safety and reducing damage from these catastrophic weather events.
