Forecasts of the July 2020 Kyushu Heavy Rain Using a 1000-Member Ensemble Kalman Filter

Abstract

Forecast performances of the July 2020 Kyushu heavy rain have been revisited with the aim of improving the forecasts for this event. While the Japan Meteorological Agency's (JMA) deterministic forecasts were relatively good, the JMA's ensemble forecasts somehow missed this event. Our approach is to introduce flow-dependence into assimilation by running a 1000-member local ensemble transform Kalman filter (LETKF1000) to extract more information from observations and to better quantify forecast uncertainties. To save computational costs, vertical localization is removed in running LETKF1000. Qualitative and quantitative verifications show that the LETKF1000 forecasts outperform the operational forecasts both in deterministic and probabilistic forecasts.

Rather than a trick to save computational costs, removal of vertical localization is shown to be the main contribution to the outperformance of LETKF1000. If vertical localization is removed, forecasts with similar performances can be obtained with 100 ensemble members. We hypothesize that running ensemble Kalman filters with around 1000 ensemble members is more effective if vertical localization is removed at the same time. Since this study examines only one case, to assess benefit of removing vertical localization rigorously when the number of ensemble members is around 1000, a larger set of cases needs to be considered in future.