采用FY-3B/IRAS亮温资料进行广义变分同化研究。广义变分同化结合了经典变分同化和稳健M-估计两者的优点。区别于经典变分同化依赖于先前的质量控制并要求误差服从高斯分布，把M-估计法耦合到经典变分同化框架中，得到广义变分同化，其弱化了同化前的质量控制和误差服从高斯分布这两个条件。目标能量泛函包含M-估计以保证对离群值具有稳健性，从而能够得到较好的同化结果。对比变分同化前后的FNL资料湿度与GDAS湿度相关系数作为同化结果检验评价。具体操作过程在FNL作为背景场的基础上分别采用经典和M-估计不同的权重因子变分同化FY3B/IRAS资料，把得到的分析场与GDAS进行相关性比较，由于湿度具有较强的非高斯性，文中首先评估了安徽省13个站GPS/PWV和积分相关湿度廓线得到大气可降水量（即GDAS/PWV和FNL/PWV资料）的相关性，进一步基于信息熵自由度思想进行了近一个月IRAS 20个通道对分析场的影响贡献率诊断研究。

This study adopts the Infrared Atmospheric Sounder of Feng-Yun and the 3rd(B) Weather Satellite(FY-3B/IRAS) brightness temperature to investigate the generalized variational assimilation, which combines the advantages of classical variational assimilation and robust M-estimators.
Classical variational assimilation is based on the model variables and satellite observations of the brightness temperature to yield a quadratic functional minimization. The observational errors are needed to follow a Gaussian distribution and subsequently apply the least-square principle. The least-squares method is sensitive to outliers; if the analyzed data contain gross errors, the parameter estimation will be inaccurate. The classical variational assimilation consists of two stages. First, an appropriate algorithm is used to identify and address outliers in the data and then, the assimilation. This approach may result in the loss of useful data because the outliers are not always harmful; some outliers may represent new information, such as weather phenomena. At present, the quality control is generally based on a certain threshold value if the subjective uncertainty is too strong. If outliers persist after the quality control, the optimal parametric results that are obtained through classical variational assimilation become meaningless.
M-estimators are added to the framework of classical variational assimilation to obtain a generalized variational assimilation, which is coupled with quality control in the process of assimilation.
The main idea is to use the weight factor of M-estimators to re-estimate the contribution rate of the observation items to the objective function in each process of objective function minimization based on classical variational assimilation. The cost function consists of M-estimators to guarantee the robustness to outliers. Thus, the assimilation results are improved. Humidity is an important dynamic variable in the NWP model. It does not only determine the occurrence of precipitation but also changes the temperature through evaporation and condensation, and it also influences the wind field by changing the pressure gradient. In addition, the nonlinearity of humidity is stronger than temperature, which causes the humidity to follow a stronger non-Gaussian distribution. Thus, humidity was used as an assimilation experiment effect validation, and the correlation coefficient of humidity was compared with FNL and GDAS, which are assimilated by different M-estimator weights. The specific operation process that is based on the FNL as the background field adopts classical and different weight factors of M-estimators to the variational assimilation of FY-3B/IRAS. In addition, the correlation between the analyzed field, and the GDAS is compared.
The correlation between 13 GPS/PWV stations of Anhui province and the integral humidity profile of the relevant field from both GDAS/PWV and FNL/PWV is evaluated. Furthermore, based on the information entropy of freedom degrees, the contribution of IRAS 20 channels were determined to analyze the field for nearly a month. Preliminary results demonstrate the potential application value of the generalized variational assimilation.