Assimilation of GPS Refractivity from FORMOSAT-3/COSMIC Using a Nonlocal Operator with WRF 3DVAR and Its Impact on the Prediction of a Typhoon Event

Abstract

A nonlocal observation operator has been developed to assimilate GPS radio occultation (RO) refractivity with WRF 3DVAR. For simplicity, in the past GPS RO refractivity was often assimilated using a local observation operator with the assumption that the GPS RO observation was representative of a model local point. Such an operator did not take into account the effects of horizontal inhomogeneity on the derived GPS RO refractivity. In order to more accurately model the observables, Sokolovskiy et al. (2005a) developed a nonlocal observation operator; which would take into account the effects of horizontal inhomogeneity on GPS RO measurements. This nonlocal observation operator calculates the integrated amount of the model refractivity along the ray paths centered at the perigee points. For comparative purposes, the nonlocal observation operator can be simplified by limiting the length of integration near the RO point. This is called the “local operator variant”, which is equivalent to the original local operator except that the original one is performed with fixed tangent points at observation levels. For computational efficiency, assimilation using both the nonlocal operator and local operator variant now is performed with smear tangent points at the mean height of each model vertical level. In this study, the statistics of observation errors using both local and nonlocal operators were estimated based on WRF simulations. The observation errors produced by the nonlocal operator are about two times smaller than those generated by the local operator and in agreement with Sokolovskiy et al. (2005b). Each of the three operators is used to assimilate GPS RO refractivity soundings from the FORMOSAT-3/COSMIC mission using the WRF 3DVAR system. The WRF model then is applied to simulate Typhoons Kaemi (July 2006) which struck Taiwan with significant torrential rainfall. The analysis increments produced by the nonlocal operator and local operator variant are quite similar in horizontal and vertical distributions; whereas, the former is slightly stretched along the ray's direction, as a result of the longer integration length. The simulated typhoon tracks prior to landfall are quite similar for the three operators. Both the nonlocal operator and local operator variant improve the detoured track after landfall as predicted by the local operator. The nonlocal operator outperforms the two local operators in rainfall prediction at later times. The performances of the nonlocal operator in general are promising and can replace the local operator at no marked cost of computational efficiency.