There
were a number of inherent errors in the models. For numerical models such as
IDW and Cokriging, the major assumption lies with how representative the data
is. It is possible that the data does not represent actual precipitation regimes
well. This is especially true in complex terrains where more than one factor
influences the pattern of precipitation (Daly et al., 1994). Furthermore, a
sparse rain gauge network can also contribute to unrepresentative data. It has
been well established that numerical models are sensitive to the number of interpolation
points (Daly et al., 1994; Prudhomme and Reed, 1999; Boer et al., 2001). The
interpolation methods may rely on elevation, but this fact may be blurred by
the effects of steepness, proximity to moisture, or the measurement site’s
exposure to prevailing winds (Whiteman, 2000). The Linear Regression model was
based on topography and its linear relationship with precipitation. It does
not consider a variety of the other factors. It is possible that a number of
the factors mentioned would be present in the GVRD. The PRISM model does consider
these factors by incorporating the concept of “topographical facets”,
where each of the facets has a unique orographic regime. The facets are best
shown with a DEM that has a resolution which closely matches the smallest orographic
scale (Daly et al., 1994).
Despite what was been discussed previously regarding the merit of cross-validation
and the assumptions associated with it, there are still a number of potential
errors. In particular, the distribution of data points and the very nature of
the interpolation methods chosen can also be a source of inaccuracy. The numerical
based interpolation methods, such as Cokriging and IDW, largely depend on the
number of data points and their distribution. Their performances tend to suffer
from a lack of points in certain regions of the study area. This is illustrated
by the northeastern region of the GVRD where there are no data points. The irregular
nature of the precipitation isohyets is a direct reflection of the lack of data
points in this region. This leads to largely inaccurate predictions in those
areas of north eastern GVRD. The regression model performs better in those areas
because it is only dependent on elevation. However, it still suffers from its
own nature of global operation. The regression model can be seen as a global
operation because it does not consider neighbouring data points in deciding
the interpolated grid value. In the overall scheme of things, the Linear Regression
model eventually suffers from the lack of consideration of neighbourhood data
points. In those areas with higher density and an even distribution of data
points, IDW and Cokriging perform better. This illustrates the importance of
neighbourhood operations and the principle of spatial autocorrelation.
