A mesoscale hybrid data assimilation system based on the JMA nonhydrostatic model

Improving the forecast skill of tropical cyclones (TCs) is important in terms of the prevention and mitigation of related disasters. To predict a hazardous event accurately, the construction of background error covariances should be more sophisticated because the analysis increment obtained through data assimilation depends on the specification of background error covariances. Thus, we developed two types of a hybrid mesoscale 4D-Var data assimilation system that uses background error covariances constructed from the perturbations in the LETKF data assimilation system. One is the spatial localization approach that assumes the correlation with the spatially separated points becomes closer to zero (4D-Var-BenkfL), while the other is the neighboring ensemble approach that regards a horizontally shifted result as a different realization of ensemble run (4D-Var-BenkfN). In order to evaluate the potential of the current-state-of-the-art 4D-Var and LETKF systems built upon the JMA non-hydrostatic model, we first conducted a single-observation assimilation experiment around a TC, followed by real-data assimilation with forecast experiments of TCs from a deterministic point of view. The single-observation assimilation experiment showed that the analysis increment obtained from the hybrid systems at the beginning of the assimilation window is the physically consistent in that, as a response to increase of TC central sea-level pressure, a cyclonic circulation in the wind field becomes weak and temperature increment has a vertically deep structure. In contrast, the analysis increment obtained from 4D-Var using merely NMC-based background error covariances (4D-Var-Bnmc) does not capture these TC related features at the beginning of the assimilation window. At the end of the assimilation window, analysis increment becomes closer to each other after the 3-h time integration with the use of high-resolution model. Nevertheless, the structure of analysis increment is relatively large in the 4D-Var using NMC-based background error covariances compared with the hybrid data assimilation systems. Analysis increment obtained from LETKF can capture the TC related features. However, there are some differences, for example, the loss of vertical coherence due to the vertical localization. Real-data assimilation experiments showed that the hybrid-based initial conditions are better than the 4D-Var-Bnmc-based initial conditions in terms of TC track prediction and intensity forecast. LETKF-based initial conditions yield the better track forecast relative to 4D-Var-Bnmc, while they do not improve the TC intensity forecast. These results are generally significant in a statistical sense. The composite analysis of steering flow around TCs indicates the resemblance of steering flow among LETKF- and hybrid-based forecasts consistent with their track forecast skill. The radius of maximum wind in 4D-Var-Bnmc-based forecasts tends to become larger as expected from the single-observation experiment, which induces too much weakening of TC vortex considering the conservation of absolute angular momentum. It contributes to the larger intensity forecast error in 4D-Var-Bnmc. LETKF generates a relatively larger vortex with higher minimum sea level pressure in the analysis and is not as good as hybrid data assimilation systems in terms of TC intensity forecast.