Usage Of Flow-dependent Background Error Covariance In Data Assimilation And Radar Wind Retrieval

Due to insufficient observation information, the quality of data assimilation depends on the background (forecast field) and it's error information crucially. The error covariance which accuracy affects data assimilation result heavily reflects the statistic structure of background error. Traditional modeled forecast error covariance which still has large difference from reality, especially for mesoscale model, is always set as temporary steady, at least in a season, spatially homogeneous and isotropic. The flow-dependent background error covariance could give more accurate background error statistic structure. This theory gives a new potential for improving data assimilation or data retrieval. In this dissertation, the usage and improvement of flow-dependent background error covariance by ensemble forecast is studied with Ensemble Square Root Filter (EnSRF),4 Dimensional Variational data assimilation (4DVAR) and radar wind retireval.The difference between flow-dependent background error covariance in EnSRF and homogeneous and isotropic background error covariance in 4DVAR is analyzed in Observing System Simulation Experiments (OSSEs) with shallow water model. The parameters for EnSRF and 4DVAR are given in this chapter. A hybrid background error covariance method for sequential assimilation is demonstrated to improve the significant estimating inaccurate of background error covariance caused by small ensemble or model bias. In the hybrid method, the steady or quasi-steady Gaussian background error covariance and the flow-dependent background error covariance by ensemble forecast are weighted together. Results show it could improve the analysis without extra computation cost. The relationship between best weighting coefficient and model error, ensemble size, and observation density is analyzed by OSSEs. A 4 Dimensional Ensemble Square Root Filter (4DEnSRF) is developed to analyze multi-time observations during an assimilation cycle.4DEnSRF will obtain better analysis, especially for unobserved model variable, when observation information isn't enough at one observation periods. Improvement isn't significant if observation information is adequate. Analysis show current localization method is suitable for 4DEnSRF. Excepting spatial localization, temporary localization also needed to be considered. The inflation factor should have spatial structure because multi-observations will be assimilated on a grid point. We researches the usage of flow-dependent background error covariance in 4 Dimensional Variational data assimilation (4DVAR) and proposes two methods, En4DVAR-V1 and En4DVAR-V2. En4DVAR-V1 calculates the background error covariance from forecasting ensemble explicitly, so localization could be applied easily. The analysis is improved clearly even with small ensemble size. But it is hard using really at present for the huge computation cost. En4DVAR-V2 preconditions the control variable by the standardized perturbed ensemble. Compared with En4DVAR-V1, it could reduce both the condition number and computation cost., Analysis error is decreased furtherly by adopting time sampling method to increase the ensemble size. Because localization isn't been used, En4DVAR-V2 performs poor then 4DVAR with small ensemble. The Ensemble Simple Adjoint method (EnSA) with flow-dependent background error covariance is developed for Doppler radar wind retrieval. This method is based on En4DVAR-V2 method. A group of perturbed equations simplified from wind advection equations are used for ensemble forecast. The correlation structure of the perturbed ensemble could reflect the variation of flow field. EnSA method obtains similar retrieval result as Simple Adjoint (SA) method. The non-localization is still affecting the retrieval.