Improving precipitation nowcasting using a three-dimensional convolutional neural network model from Multi Parameter Phased Array Weather Radar observations

In this paper, a three-dimensional convolutional neural network model (3DCNN) is proposed to nowcast a shortlived, local convective storm event by using unique 3-D observations of Multi-Parameter Phased Array Weather Radar (MP-PAWR) over Tokyo, Japan on 1 August 2019. Using statistics and forecast skill scores, nowcast skills of 3DCNN were examined with those of a three-dimensional advection nowcast model (3DNOW) which generates extrapolation-based forecasts with lead time up to 10 min. In analyzing the skill scores, two groups of a total rain area and convective rain area were made by different radar reflectivity (Z) thresholds of 10 dBZ and 37.5 dBZ, respectively. For the total rain area, it is found that 3DCNN outperformed both the 3DNOW and persistence forecast, showing the higher threat scores for all lead times. For the convective rain area, the 3DCNN and 3DNOW's performances were similar at early lead times, showing almost the same threat scores. However, the threat score of 3DCNN dropped lower than that of 3DNOW at a lead time of 10 min, indicating that 3DNOW has the better skill at relatively long lead times. Nowcasts of 3DNOW showed a limitation to yield a larger saturated Z area related to increased errors in advection vectors at longer lead times although this led to increasing the threat score.

Automated summary

This study proposed a 3DCNN model for short-term forecasting of convective storm events, showing better performance in total rainfall area prediction compared to 3DNOW. Additionally, 3DNOW demonstrated better skill at longer lead times for convective rain areas.