6. Discussion
GPS positioning results confirm that the introduction of raytraced
troposphere slant delay corrections based on numerical
weather models significantly improves results and removes
virtual station shifts caused by un-modeled atmospheric effects.
The improvement obtained from such models is strongly
pronounced during severe weather phenomena, demonstrated
here during a typhoon passage. Fine-mesh numerical weather
models have been proven to provide more realistic corrections, as
they are also capable to extract the small-scale asymmetric delay
structures which are caused by local wet refractivity fields
(i.e. clouds). Thereby, numerical weather models are not required
to provide corrections with (absolute) mm-accuracy, but correctly
predict the temporal and spatial variations. Model biases can be
estimated by a simple mathematical model within the geodetic
adjustment process and gradient approaches can be completely
turned off when ray-traced corrections are applied to the
observations. As discussed by Hatanaka et al. (2001a, 2001b)
detailed information about antenna phase center variations
(PCVs) is crucial for highly accurate positioning results. Thus,
the only disadvantage that arises from our processing strategy
stems from the fact that instrumental errors, like unmodeled
PCVs, which have azimuthal dependency are not absorbed in the
gradient model anymore. Nevertheless, this small drawback can
be overcome since instrumental corrections and antenna calibrations
are expected to get more accurate in the close future
(Schmid et al., 2007) using absolute calibration techniques. Fine-
mesh models, as those used for this study, have the advantage
over larger scale models that the topography is finer resolved and
more sophisticated physical models are used to predict the
complex weather phenomena. Therefore a better modeling of
site-specific atmospheric effects can be achieved. This permits the
use of a more robust lower elevation mask within the GPS
processing and enables the geodetic inversion to more effectively
isolate ZTD from the vertical component of position. Since such
models need to be initialized by surrounding (e.g. meso-scale)
numerical weather models it should be aimed not to introduce
large biases during this step. Moreover one has to consider that
although the model area appears to be large on the first glance, a
large fraction of the original spatial domain cannot be used in
order to treat rays at low elevation angles, too. Although a nesting
strategy has been developed (see Section 2.1), nearly half of the
model space had to be excluded for ray-tracing calculations. Thus,
the usage of fine-mesh numerical weather models is currently
only feasible for applications of regional extent.