Comparison of Hybrid 4DEnVar and 3DEnVar for the Prediction of Atmospheric Rivers

Atmospheric rivers (ARs) are elongated corridors of enhanced moisture transport from low latitudes to high latitudes. ARs play a crucial role in the global water cycle and are closely associated with weather and climate extremes in the related regions globally. Landfalling ARs can produce up to 50% of the annual precipitation for the Western United States (U.S.) and are cause of the major flooding events. Accurate forecasts of landfalling AR events can improve water management decisions and reduce the risk of flooding. However, the forecasts of ARs, in particular, the landfall events, are challenging in numerical models. There are three limiting factors for the landfall AR forecast over the Western U.S.: 1) the relatively sparse observations upstream over the Pacific Ocean; 2) the imperfect model physics and dynamics in resolving water vapor budget and vapor transport; 3) a lack of effective data assimilation methods tuned for representing the characteristics of an AR such as its filament features and the connection with low-latitude moisture source. Of them, the optimized data assimilation strategy is of paramount importance because it can take the best advantage of the available upstream observations and compensate the model errors by correcting the model towards observations. This study compares two data assimilation approaches, the hybrid 3D-EnVar and 4D-EnVar, for the forecast of four different AR cases in the Weather Research and Forecasting (WRF) model. The advantages and disadvantages of each method in assimilating the major data types over the ocean, e.g., the atmospheric wind vector (AMV) and the unique dropsonde observations taken during AR Reconnaissance (Recon), will be investigated.

Ten AR storms that AR Recon sampled during 2016, 2018, and 2019 were selected as representative AR cases for this study. For each case, four experiments were conducted using the WRF model with the Gridpoint Statistical Interpolation (GSI) system. The first experiment, called ALL_hyb3d, assimilated the AR Recon dropsonde data and other conventional data using hybrid 3DEnVar. The second experiment, called ALL_hyb4d, assimilated the same data as in ALL_hyb3d, but used hybrid 4DEnVar. The third experiment, called NoDROP_hyb3d, assimilated the same data as in ALL_hyb3d but excluded the Recon dropsonde data. The fourth and final experiment, called NoDROP_hyb4d, was the same as NoDROP_hyb3d, but used hybrid 4DEnVar. Comparisons of the four runs demonstrate that the initial analysis in All_hyb4d has the least errors in integrated vapor transport (IVT) among the four runs when compared with the high-resolution ECMWF operational analysis. On average, hybrid 4DEnVar is better at assimilating the moisture field and the meridional wind field than that hybrid 3DEnVar. However, the impacts of these two methods on model analyses are sensitive to the evolution speed and flow patterns of different AR storms. Hybrid 4DEnVar shows more advantages for more elongated ARs and systems that evolve rapidly within the assimilation window. The forecast validation shows that the All_hyb4d and All_hyb3d runs can effectively reduce the IVT errors during 6h-48h forecast than the other two sets of experiments, suggesting that runs with dropsonde data can significantly reduce moisture transport errors in an AR. The benefits of hybrid 4DEnVar methods on the forecast of ARs and precipitation skills are maximized during the 6–48 h compared to the hybrid 3DEnVar method. The average skill of 4DEnVar method is slightly higher during 48–120 h than that of 3DEnVar. However, hybrid 3DEnVar showed higher skill in terms of both bias and RMSE of integrated vapor transport during 24–72 h than hybrid 4DEnVar in two out of ten cases. This study seeks to answer the question of how different data assimilation approaches can impact AR initial conditions and forecasts, contributing to better understanding of the landfall AR physics and dynamics, and improving its associated precipitation forecast in the WRF model.