A Deep Learning Model for Precipitation Nowcasting Using Multiple Optical Flow Algorithms

• DOI: 10.1175/waf-d-23-0104.1
• Publication Date: 2023-11-14T13:13:06Z
• Authors: Ji-Hoon Ha; Hyesook Lee
• Keywords: Nowcasting; Optical Flow

Abstract The optical flow technique has advantages in motion tracking and long been employed precipitation nowcasting to track the of fields using ground radar datasets. However, performance forecast time scale models based on are limited. Here, we present results application deep learning method estimation extend its enhance nowcasting. It is shown that a model can better capture both multispatial multitemporal motions events compared with traditional methods. comprises two components: 1) regression process multiple algorithms, which more accurately captures features single algorithm; 2) U-Net-based network trains movement. We evaluated cases South Korea. In particular, minimizes errors by combining algorithms gradient descent outperforms other only algorithm up 3-h lead time. Additionally, U-Net plays crucial role capturing nonlinear cannot be captured simple advection through estimation. Consequently, suggest proposed could play significant improving current operational models, Significance Statement purpose this study improve accuracy short-term rainfall prediction methods have for By utilizing open-source libraries, such as OpenCV, commonly applied machine techniques, linear networks, propose an accessible enhancing accuracy. expect improvement will significantly practical