[M-GI29] Data assimilation: A fundamental approach in geosciences

[MGI29-01] Sparse sensor placement applied to numerical weather prediction: Experiments with a simplified AGCM

★Invited Papers

*Mao Ouyang¹, Shunji Kotsuki¹ (1.Chiba University)

Keywords:Data assimilation, Sparse sensor placement, Numerical weather prediction, Optimization, Localization

Data assimilation (DA) plays an important role in numerical weather prediction (NWP) to provide optimal initial conditions by combining forecasted state and observation data. While DA has been used to evaluate impacts of assimilated observations, limited studies tried to optimize observing placement of additional mobile observations such as by aircrafts. In this study, we applied data-driven sparse sensor placement (SSP) to optimize locations of additional weather stations. Here we consider two objective functions: maximizing the determinant and minimizing the trace of the inverse of the error covariance matrix (D- and A-optimality). We applied this two objective-function-based SSP for observing system simulation experiments using an intermediate global atmospheric model coupled with the local ensemble Kalman filter (a.k.a. SPEEDY-LETKF) to optimize observing placement in Pacific Ocean. Sensitivity analysis is investigated on the target new weather stations selection ranges with localization scales equal to 300, 500, 600, 700 and 900 km. The global average root mean square errors (RMSE) obtained from the SSP methods are compared with the ensemble spread-based determination of the observing placement. Due to the different additional station selection patterns by SSP and ensemble spread-based method, D- and A-optimality-based experiments show different global average RMSEs with the ensemble spread-based case. Examination of observing placement patterns by SSP reveals that the new stations tend to be selected near the edge of the target selection range. Our results demonstrate that the performance of SSP methods is sensitive to the localization scales of target selection range. When the localization scale of target selection range is greater than equal to 300 km and less than 600 km, A-optimality-based result shows the smallest RMSE among the other methods. When the localization scales of target selection range is greater than equal to 600 km and less than 900 km, D-optimality-based result presents the smallest RMSE. This research suggests that the data driven SSP methods could be employed to provide better observing placement approach for NWP than a simple ensemble spread-based placement.