Using a Neural Network to Predict Future Radar Frames

Nowcasting rainfall and convective weather in the 0 to 2 hour time frame continues to be a global forecasting challenge. Improved predictions of convection would be of huge benefit to sectors such as aviation, enabling airspace to be managed more safely and efficiently, such as by air traffic control pre-emptively applying flow regulations in sectors known to be impacted and giving airlines more time to find suitable alternative routes in pre-tactical planning.

Machine learning offers huge potential in this area due to the continual growth in the amount of data and computational resource available. In nowcasting, machine learning methods could add value to the more traditional physically-based extrapolation of features by:

- Calculating feature tracks and tendencies, including speed and direction, but also size, rate of growth, rate of intensification and so on.

- Allowing integration of many data sources, such as ground-based radar, satellite products, surface observations, and guidance from the Numerical Weather Prediction (NWP) model.

- Not being constrained by our incomplete knowledge of the relationships between all the chaotic structures in a given weather system.

However, our physical understanding will still be able to inform the solution, for instance in the selection of datasets which we expect to be of most importance.

This work examines the initial application of a neural network in predicting future sequences of radar imagery and methods by which to improve the performance of the network. The analysis uses a convolutional U-net architecture with sequences of 2D composite radar data (over the United Kingdom) as the initial inputs.

This talk will summarise the work to date, providing a brief introduction to neural networks, followed by a description of the data preparation, network architecture and training process. Results from the network testing will be presented and compared against a baseline optical flow methodology. The benefits and limitations of this initial application of machine learning to nowcasting will be discussed.