Meteorologists in the northern hemisphere have released a new, open-source dataset to aid in the detection and prediction of tornadoes. Given the working title “TorNet,” the dataset was curated by Mark Veillette and James Kurdzo and includes radar data from thousands of US tornadoes over the past decade. Along with the dataset, models trained on the material have also been released, potentially offering a big leap forward for storm forecasters. Tornadoes are difficult to detect and understand, despite causing significant destruction and loss of life each year. Traditionally, storm forecasters have erred on the side of caution when issuing tornado warnings, resulting in a high rate of false alarms. The TorNet dataset, which includes over 200,000 radar images, some showing tornadoes and some showing non-tornadic storms, could use machine learning to detect patterns that could eventually lead to more accurate predictions. The two researchers believe the shared use of the dataset will create a benchmark, open meteorology up to data scientists and help to solve one of nature’s enduring mysteries. Once assembled, the suite of algorithms will be tested and validated in a series of steps, enabling higher levels of operational accuracy.
Skip to content
Skip to footer
