17:15 〜 18:30
[MGI29-P02] Improving Inflation and Resampling Methods To Stabilize Local Particle Filter
キーワード：局所化粒子フィルタ、データ同化、リサンプリング、最適輸送問題、アンサンブルデータ同化
Ensemble data assimilation has been broadly applied in geoscience fields such as numerical weather prediction (NWP). While the ensemble Kalman filter (EnKF) has been intensively investigated in the past two decades, recent studies explore the local particle filter (LPF) to solve some issues in EnKF such as non-Gaussian likelihood functions. With limited particles, the LPF needs resampling that interpolates prior particles to yield posterior particles. Penny and Miyoshi (2016) developed an LPF in the form of the ensemble transform matrix of the Local Ensemble Transform Kalman Filter (LETKF). In this form, the resampling is performed by multiplying the ensemble transform matrix to the prior perturbation matrix. However, our preliminary experiments showed that the LPF was less stable than the LETKF.
This study aims at exploring better inflation and resampling methods for the LPFs that use the ensemble transform matrix for the resampling. Using the 40-variable Lorenz-96 model, we performed a series of LPF experiments with multiple inflation and resampling methods. With the simplest resampling method known as the stochastic uniform method (SU), the additive inflation was superior to relaxation to prior spread (RTPS) and additive noise to ensemble transform matrix in accuracy and stability. Following Reich (2013), we also implemented the optimal transport (OT) for generating resampling matrices. While the OT requires much more computations (O(m^1.92)) than SU (O(m)), the OT resulted in smaller root mean square errors than the SU. With the OT-based resampling, the RTPS resulted in a similar accuracy of the additive inflation. Since the incorporation of the additive inflation for the NWP system is not trivial, the combination of OT and RTPS would be a reasonable choice. This poster includes the most recent progress up to the presentation.
This study aims at exploring better inflation and resampling methods for the LPFs that use the ensemble transform matrix for the resampling. Using the 40-variable Lorenz-96 model, we performed a series of LPF experiments with multiple inflation and resampling methods. With the simplest resampling method known as the stochastic uniform method (SU), the additive inflation was superior to relaxation to prior spread (RTPS) and additive noise to ensemble transform matrix in accuracy and stability. Following Reich (2013), we also implemented the optimal transport (OT) for generating resampling matrices. While the OT requires much more computations (O(m^1.92)) than SU (O(m)), the OT resulted in smaller root mean square errors than the SU. With the OT-based resampling, the RTPS resulted in a similar accuracy of the additive inflation. Since the incorporation of the additive inflation for the NWP system is not trivial, the combination of OT and RTPS would be a reasonable choice. This poster includes the most recent progress up to the presentation.