Takemasa Miyoshi

4‐D‐Var or ensemble Kalman filter?

Abstract We consider the relative advantages of two advanced data assimilation systems, 4-D-Var and ensemble Kalman filter (EnKF), currently in use or under consideration for operational implementation.With the Lorenz model, we explore the impact of tuning assimilation parameters such as the assimilation window length and background error covariance in 4-D-Var, variance inflation in EnKF, and the effect of model errors and reduced observation coverage. For short assimilation windows EnKF gives more accurate analyses. Both systems reach similar levels of accuracy if long windows are used for 4-D-Var. For infrequent observations, when ensemble perturbations grow non-linearly and become non-Gaussian, 4-D-Var attains lower errors than EnKF. If the model is imperfect, the 4-D-Var with long windows requires weak constraint. Similar results are obtained with a quasi-geostrophic channel model. EnKF experiments made with the primitive equations SPEEDY model provide comparisons with 3-D-Var and guidance on model error and ‘observation localization’. Results obtained using operational models and both simulated and real observations indicate that currently EnKF is becoming competitive with 4-D-Var, and that the experience acquired with each of these methods can be used to improve the other. A table summarizes the pros and cons of the two methods.

Local Ensemble Transform Kalman Filtering with an AGCM at a T159/L48 Resolution

A local ensemble transform Kalman filter (LETKF) is developed and assessed with the AGCM for the Earth Simulator at a T159 horizontal and 48-level vertical resolution (T159/L48), corresponding to a grid of 480 240 48. Following the description of the LETKF implementation, perfect model Observing Systems Simulation Experiments (OSSEs) with two kinds of observing networks and an experiment with real observations are performed. First, a regular observing network with approximately 1% observational coverage of the system dimension is applied to investigate computational efficiency and sensitivities with the ensemble size (up to 1000) and localization scale. A 10-member ensemble is large enough to prevent filter divergence. Using 20 or more members significantly stabilizes the filter, with the analysis errors less than half as large as the observation errors. There is nonnegligible dependence on the localization scale; tuning is suggested for a chosen ensemble size. The sensitivities of analysis accuracies and timing on the localization parameters are investigated systematically. A computational parallelizing ratio as large as 99.99% is achieved. Timing per analysis is less than 4 min on the Earth Simulator, peak performance of 64 GFlops per computational node, provided that the same number of nodes as the ensemble size is used, and the ensemble size is less than 80. In the other set of OSSEs, the ensemble size is fixed to 40, and the real observational errors and locations are adapted from the Japan Meteorological Agency’s (JMA’s) operational numerical weather prediction system. The analysis errors are as small as 0.5 hPa, 2.0 m s 1 , and 1.0 K in major areas for sea level pressure, zonal and meridional winds, and temperature, respectively. Larger errors are observed in data-poor regions. The ensemble spreads capture the actual error structures, generally representing the observing network. However, the spreads are larger than the actual errors in the Southern Hemisphere; the opposite is true in the Tropics, which suggests the spatial dependence of the optimal covariance inflation. Finally, real observations are assimilated. The analysis fields look almost identical to the JMA operational analysis; 48-h forecast experiments initiated from the LETKF analysis, JMA operational analysis, and NCEP–NCAR reanalysis are performed, and the forecasts are compared with their own analyses. The 48-h forecast verifications suggest a similar level of accuracy when comparing LETKF to the operational systems. Overall, LETKF shows encouraging results in this study.

The Gaussian Approach to Adaptive Covariance Inflation and Its Implementation with the Local Ensemble Transform Kalman Filter

AbstractIn ensemble Kalman filters, the underestimation of forecast error variance due to limited ensemble size and other sources of imperfection is commonly treated by empirical covariance inflation. To avoid manual optimization of multiplicative inflation parameters, previous studies proposed adaptive inflation approaches using observations. Anderson applied Bayesian estimation theory to the probability density function of inflation parameters. Alternatively, Li et al. used the innovation statistics of Desroziers et al. and applied a Kalman filter analysis update to the inflation parameters based on the Gaussian assumption. In this study, Li et al.’s Gaussian approach is advanced to include the variance of the estimated inflation as derived from the central limit theorem. It is shown that the Gaussian approach is an accurate approximation of Anderson’s general Bayesian approach. An advanced implementation of the Gaussian approach with the local ensemble transform Kalman filter is proposed, where the ada...

The Non-hydrostatic Icosahedral Atmospheric Model: description and development

This article reviews the development of a global non-hydrostatic model, focusing on the pioneering research of the Non-hydrostatic Icosahedral Atmospheric Model (NICAM). Very high resolution global atmospheric circulation simulations with horizontal mesh spacing of approximately O (km) were conducted using recently developed supercomputers. These types of simulations were conducted with a specifically designed atmospheric global model based on a quasi-uniform grid mesh structure and a non-hydrostatic equation system. This review describes the development of each dynamical and physical component of NICAM, the assimilation strategy and its related models, and provides a scientific overview of NICAM studies conducted to date.

Balance and Ensemble Kalman Filter Localization Techniques

Abstract In ensemble Kalman filter (EnKF) data assimilation, localization modifies the error covariance matrices to suppress the influence of distant observations, removing spurious long-distance correlations. In addition to allowing efficient parallel implementation, this takes advantage of the atmosphere’s lower dimensionality in local regions. There are two primary methods for localization. In B localization, the background error covariance matrix elements are reduced by a Schur product so that correlations between grid points that are far apart are removed. In R localization, the observation error covariance matrix is multiplied by a distance-dependent function, so that far away observations are considered to have infinite error. Successful numerical weather prediction depends upon well-balanced initial conditions to avoid spurious propagation of inertial-gravity waves. Previous studies note that B localization can disrupt the relationship between the height gradient and the wind speed of the analysis...

Ensemble Kalman Filter and 4D-Var Intercomparison with the Japanese Operational Global Analysis and Prediction System

Abstract The local ensemble transform Kalman filter (LETKF) is implemented and assessed with the experimental operational system at the Japanese Meteorological Agency (JMA). This paper describes the details of the LETKF system and verification of deterministic forecast skill. JMA has been operating a four-dimensional variational data assimilation (4D-Var) system for global numerical weather prediction since 2005. The main purpose of this study is to make a reasonable comparison between the LETKF and the operational 4D-Var. Several forecast–analysis cycle experiments are performed to find sensitivity to the parameters of the LETKF. The difference between additive and multiplicative error covariance inflation schemes is investigated. Moreover, an adaptive bias correction method for satellite radiance observations is proposed and implemented, so that the LETKF is equipped with functionality similar to the variational bias correction used in the operational 4D-Var. Finally, the LETKF is compared with the oper...

The Local Ensemble Transform Kalman Filter with the Weather Research and Forecasting Model: Experiments with Real Observations

The local ensemble transform Kalman filter (LETKF) is implemented with the Weather Research and Forecasting (WRF) model, and real observations are assimilated to assess the newly-developed WRF-LETKF system. The WRF model is a widely-used mesoscale numerical weather prediction model, and the LETKF is an ensemble Kalman filter (EnKF) algorithm particularly efficient in parallel computer architecture. This study aims to provide the basis of future research on mesoscale data assimilation using the WRF-LETKF system, an additional testbed to the existing EnKF systems with the WRF model used in the previous studies. The particular LETKF system adopted in this study is based on the system initially developed in 2004 and has been continuously improved through theoretical studies and wide applications to many kinds of dynamical models including realistic geophysical models. Most recent and important improvements include an adaptive covariance inflation scheme which considers the spatial and temporal inhomogeneity of inflation parameters. Experiments show that the LETKF successfully assimilates real observations and that adaptive inflation is advantageous. Additional experiments with various ensemble sizes show that using more ensemble members improves the analyses consistently.

Accounting for Model Errors in Ensemble Data Assimilation

This study addresses the issue of model errors with the ensemble Kalman filter. Observations generated from the NCEP‐NCAR reanalysis fields are assimilated into a low-resolution AGCM. Without an effort to account for model errors, the performance of the local ensemble transform Kalman filter (LETKF) is seriously degraded when compared with the perfect-model scenario. Several methods to account for model errors, including model bias and system noise, are investigated. The results suggest that the two pure bias removal methods considered [Dee and Da Silva (DdSM) and low dimensional (LDM)] are not able to beat the multiplicative or additive inflation schemes used to account for the effects of total model errors. In contrast, when the bias removal methods are augmented by additive noise representing random errors (DdSM1 and LDM1), they outperform the pure inflation schemes. Of these augmented methods, the LDM1, where the constant bias, diurnal bias, and state-dependent errors are estimated from a large sample of 6-h forecast errors, gives the best results. The advantage of the LDM1 over other methods is larger in data-sparse regions than in data-dense regions.

Effective assimilation of global precipitation: simulation experiments

Past attempts to assimilate precipitation by nudging or variational methods have succeeded in forcing the model precipitation to be close to the observed values. However, the model forecasts tend to lose their additional skill after a few forecast hours. In this study, a local ensemble transform Kalman filter (LETKF) is used to effectively assimilate precipitation by allowing ensemble members with better precipitation to receive higher weights in the analysis. In addition, two other changes in the precipitation assimilation process are found to alleviate the problems related to the non-Gaussianity of the precipitation variable: (a) transform the precipitation variable into a Gaussian distribution based on its climatological distribution (an approach that could also be used in the assimilation of other non-Gaussian observations) and (b) only assimilate precipitation at the location where at least some ensemble members have precipitation. Unlike many current approaches, both positive and zero rain observations are assimilated effectively. Observing system simulation experiments (OSSEs) are conducted using the Simplified Parametrisations, primitivE-Equation DYnamics (SPEEDY) model, a simplified but realistic general circulation model. When uniformly and globally distributed observations of precipitation are assimilated in addition to rawinsonde observations, both the analyses and the medium-range forecasts of all model variables, including precipitation, are significantly improved as compared to only assimilating rawinsonde observations. The effect of precipitation assimilation on the analyses is retained on the medium-range forecasts and is larger in the Southern Hemisphere (SH) than that in the Northern Hemisphere (NH) because the NH analyses are already made more accurate by the denser rawinsonde stations. These improvements are much reduced when only the moisture field is modified by the precipitation observations. Both the Gaussian transformation and the new observation selection criterion are shown to be beneficial to the precipitation assimilation especially in the case of larger observation errors. Assigning smaller horizontal localisation length scales for precipitation observations further improves the LETKF analysis.

Comparison of Local Ensemble Transform Kalman Filter, 3DVAR, and 4DVAR in a Quasigeostrophic Model

Local ensemble transform Kalman filter (LETKF) data assimilation, three-dimensional variational data assimilation (3DVAR), and four-dimensional variational data assimilation (4DVAR) schemes are implemented in a quasigeostrophic channel model. Their advantages and disadvantages are compared to assess their use in practical applications. LETKF and 4DVAR, which take into account the flow-dependent errors, outperform 3DVAR under a perfect model scenario. Given the same observations, LETKF produces more accurate analyses than 4DVAR with a 12-h window by effectively correcting the fast-growing errors with the flow-dependent background error covariance. Even though 4DVAR performance benefits substantially from using a longer assimilation window, LETKF is also able to achieve a satisfactory accuracy compared to the 24-h 4DVAR analyses. It is shown that the advantage of the LETKF over 3DVAR is a result of both the ensemble averaging and the information about the ‘‘errors of the day’’ provided by the ensemble. The analysis corrections at the end of the 12-h assimilation window are similar for LETKF and the 12-h window 4DVAR, and they both resemble bred vectors. At the beginning of the assimilation window, LETKF analysis corrections obtained using a no-cost smoother also resemble the corresponding bred vectors, whereas the 4DVAR corrections are significantly different with much larger horizontal scales.