Analysis of the Fraction Skill Score for the rainfall verification of high resolution model in KMA

 The Korea Peninsula is surrounded by the sea on three sides and is made up of mountainous areas more than 70% of the land, making it difficult to predict and verify precipitation using numerical models. To overcome these topographical features, a dense observation network and 10 radar site have been operated in Korea Meteorological Administration (KMA). Also, a regional prediction system based on the Korea Intergrated Model (RDAPS-KIM) has been operated since May 2022. RDAPS-KIM covers the East Asia region with a finer horizontal resolution(3km) than the global model KIM(12km) and simulates real complex terrain closely.
 Traditional skill score to verify models assesses the rainfall prediction performance for the grid closest to single point locations, which has the double penalty problem that the forecast precipitation exhibits the same pattern as the observation but leads to worse prediction evaluations due to missing spatial displacement. This double penalty issue occurs more frequently in regional model when the rainbands were shifted, so it can lead to misconception that regional model underperform global model. In order to avoid this problem, spatial verification methods to evaluate probability of rainfall forecast in the surrounding area have been suggested in the previous study (Ebert, 2008). The Fraction Skill Score (FSS) is one of the spatial verification methods suggested Roberts and Lean (2008), which allows the comparison of predicted precipitation with a spatial truth fields such as radar data. Accordingly, by verifying prediction with probability, it is expected that the shortcomings of verification of high resolution models can be overcome. Also, the FSS is recommended for verification in numerical weather prediction models by World Meteorological Organization (WMO) (JWGFVR, 2013).
 In this study, the models of KMA (KIM and RDAPS-KIM) are assessed using FSS method compared with different sized neighborhoods and various rainfall threshold for 3h-accumulated precipitation. Analysis indicated how the spatial scale influences the FSS values, showing that FSS increases as neighborhood size increases. The use of selected one proper neighborhood size pointed out that RDAPS-KIM lead to a high FSS than KIM, because it fits the rainbands better. The result shows that RDAPS-KIM is more efficient for heavy precipitation and local rainfall than KIM. In the future, we will operate the higher resolution 1km model to improve prediction performance in simulating severe weather events, and evaluate in combination with traditional techniques and probabilistic forecasts. It is expected that interpretation of accuracy of precipitation forecasts from various perspectives will be possible.