Application of Optical Flow Techniques to Rainfall Nowcasting

The Hong Kong Observatory has been operating an in-house developed nowcasting system, namely “Short-range Warning of Intense Rainstorms in Localized System (SWIRLS)”, to support the operations of rainstorm nowcasting and warning in Hong Kong, as well as to provide rainfall nowcast services directly for the public. SWIRLS works primarily by extrapolating radar echoes, in which the most critical step is to track radar echoes for the generation of their motion fields. SWIRLS adopted a correlation-based method in its first operational version in 1999, subsequently replaced by an optical flow algorithm in 2010. While outperforming the correlation-based method, the optical flow approach was found to underestimate motion speeds at times. To tackle this problem, a novel algorithm was developed and put into trial operation in 2012. This algorithm captures the most critical radar echoes and analyse their motion fields by the real-time variational optical flow computation proposed by Bruhn et al., taking advantages of the Horn and Schunck approach and the Lucas-Kanade method. This paper describes the principles of the above three radar echo tracking algorithms, examines their performances in several significant rainstorm cases, and compares their overall performances through systematic verification based on data from June 2012 to May 2014. The limitations of the algorithms and of the nowcasting system SWIRLS are discussed. The feasibility of generating probabilistic nowcasts through perturbing the parameters adopted in optical flow algorithm is explored and illustrated with an experimental ensemble rainfall nowcast system. Other future development potentials are also presented.