Strategies to Overcome Filter Divergence in the DRP-4-DVar Approach
Strategies to overcome filter divergences in DRP-4DVar approach

This paper discusses an important issue related to filter divergence in the dimension-reduced projection, four-dimensional variational data assimilation (DRP-4-DVar) approach. Idealized experiments with the Lorenz-96 model over a period of 200 days showed that the amplitudes of the root mean square errors (RMSEs) reached the same levels as those of the state variables after approximately 100 days because of the accumulation of sampling errors following the cycle of assimilation. Strategies to reduce sampling errors are critical to ensure the quality of ensemble-based assimilation. Numerical experiments showed that localization and reducing observational errors can alleviate, but cannot completely overcome, the filter divergence in the DRP-4-DVar approach, while the method of perturbing observations and the inflation technique can efficiently eliminate the filter divergence problem.