Model-based assimilation of atmospheric moisture observations is complicated by the lack of meaningful error statistics, both for the data themselves and for the background estimates provided by the model. Part of the problem is the high variability, in time and space, of actual moisture fields in the atmosphere. Since data assimilation methods essentially produce spatially smoothed averages of observations and background, it is crucial that the variable chosen to represent moisture lends itself well to averaging. This is not the case, for example, for variables that have traditionally been used for moisture analysis, including atmospheric water vapor mixing ratio and the logarithm of specific humidity.
Based on a statistical analysis of observed-minus-forecast residuals collected in different regions and different seasons, we show that a more suitable analysis variable is obtained by scaling the observed mixing ratio by the model's saturation mixing ratio. This variable is similar to relative humidity but does not present the difficulties associated with analyzing actual relative humidity data, which is an inherently multivariate problem requiring information about covariances between temperature and moisture errors. Use of the 'pseudo' relative humidity computed from the model's saturation mixing ratio, however, leads to a univariate analysis for which approximate error covariances are relatively easy to obtain.
We will show, based on statistics of observed-minus-forecast residuals, that the introduction of this new moisture variable causes a significant improvement of the quality of the moisture analysis. We will also describe our method of assimilating total precipitable water in the fvDAS, and present preliminary results that show the impact of these data on the assimilation.