Examination and evaluation of deep learning models for radar echo nowcasting in Wuhan area

Based on four deep learning models(PredRNN++、MIM、CrevNet and PhyDNet), used the radar and precipitation data in Wuhan area from 2012 to 2019 and defined the radar echo area index, we examined and evaluated the forecasting performance of the four deep learning algorithms in nowcasting of radar echo with different echo area in Wuhan. The results are summarized as follows: (1) All models'forecasting ability decreases rapidly with the increase of echo intensity. The POD (Probability Of Detection) and CSI (Critical Success Index) of normal intensity are much higher than those under the strong intensity, while the FAR (False Alarm Rate) is far lower. (2) For both the normal intensity and the strong intensity radar echo, with the increase of radar echo's area all the models'POD increases and FAR decreases. As a result, the CSI improves. But those variation amplitude is more significant under the normal intensity. (3) In all types of echo areas, for both the normal intensity and strong intensity echo radar, the CSI and POD of PredRNN++ are the highest, while those of the CrevNet's are lower. The FAR of MIM is the lowest, but the differences between the models are more obviously under the normal intensity. and those differences may be mainly caused by the different intrinsic structure between each model. (4) Regardless of the area and intensity of radar echo, with the increase of forecast time, the POD of all deep learning models decreases slowly, while the FAR increases slowly. Therefore, the CSI decreases slowly. But with the forecast time extending, both the decline and the increase are smaller. However, the difference between large area and small area increases gradually.