A Comparison between Digital Filtering Initialization and Nonlinear Normal-Mode Initialization in a Data Assimilation System

Abstract

The objective of this study is to examine the performance of the adiabatic digital filtering initialization scheme of Lynch and Huang, the diabatic digital filtering initialization scheme of Huang and Lynch, and the diabatic nonlinear normal-mode initialization scheme of Cederskov in a complete data assimilation system. In particular, the authors wish to examine the handling of observations and the changes that the initialization makes to the analysis in an intermittent data assimilation cycle. As a reference the authors use the adiabatic nonlinear normal-mode initialization of Machenhauer, formulated according to Bijlsma and Hafkenscheid, which is the current operational initialization scheme at the, Danish Meteorological Institute. The initialization schemes tested are found to produce a well-balanced model state that is at least as good as that produced by the reference scheme. Furthermore, the changes to the analysis made by the different initialization schemes are similar and the observations are therefore treated similarly with the different schemes. It is thus found that the introduction of a new initialization procedure has no detrimental effect on the data assimilation cycle. On the contrary, the two diabatic schemes reduce the noise level considerably compared to the adiabatic ones albeit at an increased computational cost. Considering the advantages of a diabatic scheme, in particular the future possibility of including cloud properties in the initialization procedure (Huang and Sundqvist), the use of a diabatic scheme seems well justified. The noise reduction is perhaps not the most important aspect as all schemes behave identically in the handling of observations. Instead, the possibility of including satellite-derived cloudiness and precipitation data in the analysis and initialization cycle is a much move important aspect. From this point of view the digital filter has a clear advantage over the normal-mode initialization scheme as all dependent variables of the model are initialized.