[A-AS02] Weather, Climate, and Environmental Science Studies using High-Performance Computing

[AAS02-05] Enhancing Small-Scale Global Weather Forecasting by High-Frequency Satellite Data Assimilation: A Horizontal Localization Aspect

*Rakesh Teja Konduru^1,2, Jianyu Liang^1,2, Shigenori Otsuka^1,2, Takemasa Miyoshi^1,2 (1.Data Assimilation Research Team, RIKEN Center for Computational Science, Kobe, Japan, 2.Prediction Science Laboratory, RIKEN Cluster for Pioneering Research, Kobe, Japan)

Keywords:High-frequency observations, Satellite Data Assimilation, Error decomposition, LETKF, Localization, NICAM

In the Local Ensemble Transform Kalman Filter (LETKF), horizontal localization (HLOC) determines the spatial impact of observations on the analysis increments. HLOC is crucial for mitigating sampling errors caused by limited ensemble size and addressing spurious correlations between distant points. Effective tuning of HLOC is essential for optimizing the localization scale, thereby improving forecast accuracy. The sensitivity of HLOC to the assimilation frequency of satellite data was not well investigated so far. In this study we explored sensitivity of HLOC to the assimilation frequency of satellite data by employing the NICAM-LETKF data assimilation system through hypothetical Observing Systems Simulation Experiments (OSSEs). We conducted four experiments, assimilating clear-sky AMSU-A satellite radiances at different frequencies: hourly (1H), bi-hourly (2H), every three hours (3H), and every six hours (6H), in addition to conventional observations. The Root Mean Square Error (RMSE) of air temperature relative to the nature was computed for each experiment since AMSU-A is sensitive to the air temperature profile. We analyzed the RMSE and air temperature analysis increments, decomposed them down into different temporal scales using the Diagnostic Scale Decomposition (DSD) method. Our results indicated that a default HLOC setting in each experiment resulted in higher RMSEs in the 3-6 hours and 12-hour high-frequency scales compared to the optimally tuned HLOC for each experiment. At high-frequency scales, the 1H assimilation with optimal HLOC outperformed the 2H, 3H, and 6H frequencies. The optimal HLOC radius was found to be smaller for the 1H and 2H frequencies than the default setting. Also, the analysis increments were reduced with the optimal smaller HLOC in the high-frequency scales for 1H and 2H. However, a very tight HLOC in 1H/2H scenarios might introduce additional imbalances, particularly in larger scales, due to the tighter localization combined with the imbalances from frequent assimilation. Addressing these challenges requires a careful understanding on the balance between localization scale and assimilation frequency. Looking ahead, an adaptive HLOC strategy specifically designed for scenarios involving frequent assimilation is a potential future direction.