Researchers at MIT Lincoln Laboratory have introduced an open-source dataset called TorNet in an attempt to enable enhanced detection and prediction of tornadoes. The dataset comprises radar returns from thousands of tornadoes that struck the US over the past decade and includes copies of storms that generated tornadoes as well as other extreme weather events that did not. The dataset contains over 200,000 radar images, with 13,587 depicting tornadoes.
The researchers hope the dataset will provide a foundation for machine learning algorithms and unveiled models trained on it. They claim these models can spot a tornado, a development which could open new frontiers in forecasting, leading to more accurate warnings and potentially saving lives.
Tornadoes are infamously hard to predict with over 70% of tornado warnings turning out to be false alarms. Many scientists have turned to machine learning to better detect and predict tornadoes, however, making sense of raw datasets has been a challenge. TorNet aims to fill this gap by including all the necessary elements such as reflectivity and radial velocity.
The team has successfully applied deep learning to process visual data, discovering key features across the dataset’s images. The results have been encouraging, with the algorithm accurately classifying tornadoes, particularly those of a higher EF rating.
To further operationalise the application of this model, the next steps would involve researchers worldwide developing their own algorithms and putting them into test beds. This would provide forecasters access to these algorithms and start the process of transitioning it into operations, potentially increasing trust in tornado predictions and reducing the rate of false alarms.
