Elevation Model Evaluation
The jackknife validation procedure was repeated to evaluate the elevation adjusted model.
The min and max temperatures at each of the 233 stations were predicted with the improved kriging model
using all data within the 2.2 degree cutoff, except for stations within 0.1 degrees.
This excludes the station itself and a few other very close stations.
The mean absolute error for this procedure was 3.27°F, or an improvement of 0.25°F.
As before, this jackknife error is a near worse-case situation.
More important than the small improvement in absolute error is the removal of all of the elevation bias
as seen by the zero slope in the graph below.
While the errors in this worse-case situation can approach +/- 50°F, it is very rare to exceed 25°F.
The 50% and 95% intervals in the graph below indicate that most errors are much smaller.
The model should now be able to deal better with locations where the elevation differs substantially
from the elevations of the surrounding air temperature stations.
For the best-case situation (using the stations own data), the mean absolute error with the
improved kriging model is 0.054°F.
This error is increased from the non-elevation corrected model by 0.03°F.
This is expected as there is no way to improve the best-case kriging result by adjusting for elevation.
The impact of the elevation correction is, however, very small because the elevation difference in this
situation is also small.
In the best-case situation, over 99% of the errors are within 0.1°F, over 99.9% are within 1.0°F
and all errors are within just 5.5°F of the observed result.
Again, the normal error encountered will be somewhere between these best-case and the worse-case situations.
27 AUG 2001, updated on 30 OCT 2001
dlg@rapid.msu.montana.edu