Tn. Krishnamurti et al., ON THE IMPROVEMENT OF PRECIPITATION FORECAST SKILL FROM PHYSICAL INITIALIZATION, Tellus. Series A, Dynamic meteorology and oceanography, 46(5), 1994, pp. 598-614
This study explores the impact of physical initialization on the numer
ical weather prediction of tropical rainfall. The goal is to improve t
he definition of the initial state by assimilation of proposed or curr
ently available surface and satellite-based observations during a pre-
integration phase using a global spectral model. Physical initializati
on refers to the use of reverse algorithms consistent with the physics
of the numerical model which can provide a modification of the initia
l state via incorporation of an analysis of tropical rainrates. This m
odification of the initial state variables is accomplished in an assim
ilation phase of the model forecast. The physical initialization proce
ss produces, in a diagnostic sense, a thermodynamic consistency betwee
n the humidity variable, the surface fluxes, rainfall distributions, d
iabatic heating and the clouds. A diabatic initialization is achieved
by a Newtonian relaxation of the above diagnosed humidity variable whe
re the divergent wind is permitted to evolve in response to the impose
d surface fluxes and the condensation heating in a consistent manner.
An important finding of this study is related to the absolute correlat
ion of the observation only based rainfall and the model-based rainfal
l at the initial time and at the end of a one day forecast which are s
ignificantly improved with the use of physical initialization. The ''o
bserved'' rainrates are obtained from algorithms that translate satell
ite-based measurements of outgoing longwave radiation and radiances fo
r an array of microwave frequencies. In addition, the available rainga
uge records over the land area are incorporated to define the ''observ
ed'' rainfall over the gaussian transform grid squares of a global spe
ctral model at a high resolution (T106). Thus the rainfall measures ar
e averages over roughly 100 x 100 km(2) and 7.5 min which is the time
step of the spectral model. It is for this averaged representation tha
t we are able to demonstrate a very marked improvement in nowcasting a
nd one day forecasts of tropical rainfall. The monthly mean rainfall c
limatology, thus obtained, nearly replicates the rainfall analyses pro
vided to the physical initialization.