A deep learning approach for probabilistic forecasts of cumulonimbus clouds from NWP data

Lightning, hail, severe turbulence and severe icing associated with cumulonimbus clouds (Cb) present a significant safety hazard to air traffic and can impact the comfort and timeliness of a flight. The World Area Forecast System (WAFS) facilitates safe and efficient flight planning by providing global forecasts of key meteorological hazards. The next generation of WAFS will provide probabilistic forecasts of these hazards, including cumulonimbus clouds.

At the Met Office, these forecasts are currently made using three simple threshold tests applied to parameters from MOGREPS-G, a global NWP ensemble. These thresholds are used as a proxy for the occurrence of cumulonimbus clouds in the NWP data.

In this work, a series of deep learning models have been trained to predict the occurrence of cumulonimbus in global satellite observations using a wider set of parameters from the control member of MOGREPS-G. The purpose of the training is for the deep learning model to learn the representation of a cumulonimbus in the NWP data in a supervised manner. The model predictions are then applied to the whole ensemble to produce a probability forecast of cumulonimbus occurrence.

A range of loss functions were used during model training and verification to account for spatial information at a range of scales. Different loss functions were also used to enhance the reward for correct forecasts of the relatively rare cumulonimbus clouds.

Some of the trained models are shown to have greater skill than a baseline using the threshold test method. The model characteristics change depending on the choice of loss function used during training.

Further work is needed to explore how to make predictions at a range of lead times and how to use inputs from the whole ensemble.