Hail nowcasting based on radar and model fields using random forests

Radar composites from MYRORSS and model reanalyses fields from ECMWF over five summers (1999-2003) are used to design a nowcasting system based on random forests to predict the likelihood of different thresholds of maximum estimated size of hail (MESH). For each MESH pixel to be predicted, 14 radar-based inputs and 11 model-based inputs are used to train the machines.

Noteworthy peculiarities to the approach include:

1) The design and testing of three distinct machines for nowcasting at three forecast horizons of 15 min, 30 min, and 60 min;

2) The comparison between machines trained with the measured or estimated data alone and machines trained with a five-ensemble dataset where one contains the data and the four others contain modified data with realizations of expected errors added to the input.

These allow us to:

1) Study the shift in the value, or lack of, between radar and model parameters with increasing forecast horizons and gauge the need for dedicated AI machines for each forecast horizon; and,

2) Evaluate the benefit of also providing data with errors to help the machine deal with the imperfect data it will likely have to face when it will be used.

Results are pending at the time of this writing.