Zx. Pu et al., FORECAST SENSITIVITY WITH DROPWINDSONDE DATA AND TARGETED OBSERVATIONS, Tellus. Series A, Dynamic meteorology and oceanography, 50(4), 1998, pp. 391-410
While forecast models and analysis schemes used in numerical weather p
rediction have become generally very successful, there is an increasin
g research interest toward improving forecast skill by adding extra ob
servations either into data sparse areas, or into regions where the ve
rifying Forecast is most sensitive to changes in the initial analysis.
The latter approach is referred to as ''targeting'' observations. In
a pioneering experiment of this type, the US Air Force launched dropwi
ndsondes over the relatively data sparse Northeast Pacific Ocean durin
g 1-10 February 1995. The focus of this study is the forecast sensitiv
ity to initial analysis differences, forced by these observations by u
sing both the adjoint method (ADJM) and quasi-inverse linear method (Q
ILM), which are both useful for determining the targeting area where t
he observations are most needed. We discuss several factors that may a
ffect the results, such as the radius of the mask for the targeted reg
ion, the basic flow and the choice of initial differences at the verif
ication time. There are some differences between the adjoint and quasi
-inverse linear sensitivity methods. With both sensitivity methods it
is possible to find areas where changes in initial conditions lead to
changes in the forecast. We find that these two methods are somewhat c
omplementary: the 48-h quasi-inverse linear sensitivity is reliable in
pinpointing the region of origin of a forecast difference. This is pa
rticularly useful for cases in which the ensemble forecast spread indi
cates a region of large uncertainty, or when a specific region require
s careful forecasts. This region can be isolated with a mask and forec
ast differences traced back reliably. Another important application fo
r the QILM is to trace back observed 48-h forecast errors. The 48-h ad
joint sensitivity, on the other hand, is useful in pointing out ar:as
that have maximum impact on the region of interest, but not necessaril
y the regions that actually led to observed differences, which are ind
icated more clearly by QILM. At 72 h, the linear assumption made in bo
th methods breaks down, nevertheless the backward integrations are sti
ll very useful for pinning down all the areas that would produce chang
es in the regions of interest (QILM) and the areas that will produce m
aximum sensitivity (ADJM). Bo:h methods can be useful for adaptive obs
ervation systems.