From
Figure 22,
it can be seen that there is a deviation between the axis of maximum variation
in annual precipitation. This is shown by rotating the station locations to
reveal the orientation of maximally contrasting variation between the perpendicular
axes. It can be seen that there exists an orientation for which there is basically
zero overall trend along one axis and a strong trend along the other. A stronger
interpolation model would incorporate these aspects. An anisotropic IDW model
could replace the conventional distance value measuring two-dimensional distance
with a variable measuring distance only in one-dimension, that is, along the
axis of the dominant trend. For example, with reference to Figure 23,
the interpolation of values at points B and D would only consider y-distance
from samples A and C rather than the conventional two-dimensional distance.
This model would then be analogous to idealized example described above –
precipitation could be predicted simply by weighting station precipitation values
according to their distance from the unsampled point as measured only along
one axis, without the explicit inclusion of elevation data. Although such a
model has not been constructed in the current study, it is expected that it
could display marked improvement over the models presented.
Anisotropic Cokriging methods would similarly be expected to produce superior
results to their isotropic counterparts for this study area. However, it is
possible that the inclusion of orographic elevation as a model input may still
not result in an improvement over the simpler, univariate IDW technique. If
not, this would simply offer further support to the concept that, although there
is a strong elevational trend, it is not a particularly useful model input in
this scenario because it is already implicitly contained within the spatial
structure of the available precipitation data. This conclusion was not expected,
but it is a fascinating concept as well as being extremely practical to be able
to show that more complex or more inclusive models will not always produce better
results.