Abstract:Based on the radar echoes open data set of Guangdong-Hong Kong-Macao Greater Bay Area, first the optical flow motion vector and the future 120 minutes extrapolation echoes are calculated by Lucas-Kanade optical flow method based on semi-lagrangian advection scheme, then they were used as input features for deep learning (SmaAt-UNet) with the past observed radar echoes. Accordingly, the applicability of optical flow motion vector and extrapolation radar echoes in improving deep learning extrapolation performance is investigated based on echoes image, Critical Success Index (CSI), Probability of Detection (POD), and False Alarm Rate (FAR). Results show that the performance of DL in extrapolating the location, intensity and evolution of the generation and extinction is significantly improved by the addition of the optical flow motion vector, but weak by the error of optical flow extrapolation echoes. Therefore, atmospheric dynamical features can provide effective atmospheric evolution information for deep learning, which may improve their nowcasting capability.