A Study on Nowcasting Using a Pyramidal Optical Flow

The optical flow estimates the movement of an object in successive images. The method has been applied as one of common methods for calculating motion vectors in nowcasting systems based on extrapolated observation of radar data (i.e., images) as research or operational purpose. This method assumes that intensity of the rainfall objects (e.g., reflectivity or rainrate in radar images) are not change with time and remains constantly. Since the solution to the equations embedded in the optical flow is represented by a linear equation, the method divides the analysis domains into smaller areas (defined as blocks, e.g., 10 X 10 pixels²) and find a single motion vector using a least squares approach for each block. When the block size is too large, small-scale motion cannot be resolved, while small block sizes are unable to calculate motion vectors larger than block sizes. It is important to determine the optimal block size because the possible size of motion vectors depends on the block size. To solve this problem, an image pyramid technique was proposed. The image pyramid is constructed of a multi-resolution structure. Moreover, the method enables to detect large and small motions with a fixed block size. This method is composed smoothing the image with an appropriate smoothing filter and its subsampling. The purpose of this study is to investigate the accuracy of motion vector for block sizes and is to evaluate the performance of rainrate simulated from the improved model forecast (image pyramid technique is applied). The Short-Term Ensemble Prediction System (STEPS) is a probabilistic nowcasting system based on the extrapolation of radar images and is developed in the Australian Bureau of Meteorology and the United Kingdom Met Office. In this study, the STPES provides real-time 15-members ensemble precipitation forecasts with 1 km spatial resolutions and 10 minutes temporal resolution up to 6 h lead time on a 1024 X 1024 km domain. The image pyramid was constructed in three spatial resolutions as 1, 2, and 4 km (i.e., 1024 X 1024, 512 X 512, and 256 X 256 pixels). The performance was examined with statistical scores using Radar-AWS Rainrate (RAR) data in May to July 2016, and an initialization is performed every 6 h. It is expected that the result of this study can be utilized as a proof-of-a-concept to determine the optimal block size.

Acknowledgment

This work was funded by the Korea Meteorological Industry Promotion Agency under Grant KMIPA 2015-5060 and the BK21 plus Project of the Graduate School of Earth Environmental Hazard System.