|
This study is about the development of a deep neural network to make very short-term predictions of torrential rains at the urban scale (meso-γ). The new polarimetric Phased Array Weather Radar (MP-PAWR) operating at Saitama (Japan) since 2018 is used. Thanks to the unique spatio-temporal resolution of the measurements, the precursors of torrential rains are detected aloft more than 20 minutes before the rain occurs. With this information, we aim at the prediction of surface precipitation with a lead time of 20 min, a horizontal resolution better than 500 m within a radius of 25 km around the instrument. Two supervised neural networks are considered to extrapolate radar reflectivity (ZH) at the altitude of 600 m. The first model (model-1) is based on a technique developed for mesoscale predictions from observations at a single altitude. It uses horizontal (2D) convolutions in gated recurrent time iterations and a multilayer encoder-decoder (EC/DC) architecture. The technique is adapted to consider 3 radar parameters and 11 altitudes up to 10 km, in the same way as RBG channels in video analysis. The second model (model-2) uses similar architecture but with 3D spatial convolutions to properly describe the vertical motions between adjacent layers. The input to the models consist in 20 min long time series of ZH, Doppler velocity and differential reflectivity observations (30 sec sampling). The models are trained using all the rain events observed between August and October 2018, and are assessed using local heavy rains observed over a period of 1-hour on July, 24, 2018. The beginning of the rain is first predicted with a lead time of about 5 min, and its evolution is fairly well reproduced to lead times up to about 10 min. Results quickly degrades for longer lead times. We found that a deeper network with 4 layers EC/DC gives better 20 min predictions than a model with 3 layers, but final results were not yet obtained at the time of writing. Regarding lead-times of 10 min, model-2 gives critical success indexes (CSIs) of 0.60 and 0.40 for pixels with ZH> 10 dBZ and 37 dBZ, which is comparable or better than results presented in other studies. For lead-times of 20 min, CSIs dropped to 0.28 and 0.10, respectively, and no other studies was found for comparison. Model-1 clearly shows poorer performance, especially for high ZH. However, this approach demands much less calculations and the training lasts only 2 weeks long, namely half of the time spent for model-2. Therefore, it is worth further studying both approaches and potential improvements are discussed.
|
