After a tornado forms, a U-Net forecasts thirty minutes of radar reflectivity at the level of the national weather model

Tornadoes cause an average of more than 50 deaths and $100 million in damage per year in the United States. The network produces its probabilistic forecasts much faster than a physical model.

When a tornado becomes probable, weather radar is the tool that informs forecasters about the morphology of the storm, its displacement and the evolution of its intensity. This work adds a deep U-Net-like learning system to predict the evolution of radar reflectivity in the very short term following the formation of a tornado — an average cost of more than 50 lives and $100 million in damage per year in the United States.

The model includes radar imaging of the Multi-Radar Multi-Sensor (MRMS) data set and near-storm environmental data from the HRRR numerical prediction model. Driven on a corpus of tornadostorms, it produces thirty minutes of probabilistic predictions, obtained by predicting the parameters of a SinhArcSinh distribution rather than a single value.

The result is double. The forecasts remain physically realistic and achieve a quality comparable to the one hour HRRR forecast, with a reasonable probabilistic calibration and a set of explanation methods for end users. Above all, they are obtained much faster than with a digital model — a decisive argument for a transition to an operational forecast room, which the authors envisage after further developments.