[M-GI26] Data assimilation: A fundamental approach in geosciences

[MGI26-P06] Incorporate ensemble sensitivity into dynamic background uncertainty to improve adaptive observation in numerical weather prediction

*欧陽 懋¹、小槻 峻司¹ (1.千葉大学)

キーワード：スパースセンシング、データ同化、天気予報

Our previous research has demonstrated that sparse sensor placement (SSP) could select optimal observation locations in the geophysical models, e.g., sea surface temperature. The optimal locations are selected based on the least square rule to minimize the static background uncertainty. This study tries to introduce SSP into the numerical weather prediction (NWP) for improving the weather forecast by adding an additional mobile observation. The key challenge is that the background uncertainty of NWP is dynamic, not static, due to the assimilation of observation data. Correspondingly, we propose two methods in this study: (1) incorporate ensemble sensitivity to represent dynamic background uncertainty; (2) replace the objective function in SSP to be the dynamic background uncertainty. Ensemble sensitivity is the linear regression of analysis errors onto the background uncertainty. The model used here is an intermediate global atmospheric model coupled with the local ensemble Kalman filter (a.k.a. SPEEDY-LETKF). The objective is to reduce one-year average global root mean square error (RMSE) through assimilating an addition mobile observation within the selection ranges (300 km ~ globe) centered at the southern Pacifica Ocean. Control case is the ensemble spread method, which selects an observation location in the largest standard deviation of ensemble members. Our results demonstrated that RMSE shows a decreasing trend when the observation selection ranges increase from 300 km to 3000 km for all cases. Incorporating ensemble sensitivity to represent dynamic background uncertainty could improve the NWP for an additional mobile observation by SSP, e.g., RMSE shows around 35 % reduction compared with the control case when selection range is 1200 km. The dynamic background uncertainty here includes the pattern of error growth, which might be beneficial in selecting the optimal additional observation location. If we replace the objective function in SSP to be the dynamic background uncertainty, the results can be further improved. This is reasonable due to that SSP can select the location with largest background uncertainty, consequently, mostly reduce the global analysis error. Our research suggests that considering the ensemble sensitivity could reasonably represent the dynamic background uncertainty, and SSP can be implemented to select adaptive observation locations accordingly.