Figure 6. Accuracy comparison map

After converting all layers to raster representation, a raster calculation was performed to subtract reference data from all above kriging results. The resulting map shows the offset of each kriging result in terms of degree Celsius.

Investigating the kriging layer, it is clear that the interpolation method maintains its accuracy layer in an area of rapid elevation change and the Delta region. However, it tends to overestimate in regions of constant elevation and underestimate in regions of high elevation; as can be seen in the northern region of the study area. The same trend can also be seen in the IDW layer where the model was accurate and inaccurate in the same regions. This trend is quite interesting to see since the exact cause of this result is unknown. Both interpolations are transformed by DALR which might be a contributing factor.

Investigating the cokriging layer, it maintains its accuracy in the coastal areas and overestimates everywhere else. There is a region of red surrounding the Delta Burns Bog climate station which is caused by the classification scheme that highlights red in areas where the offset is more than -0.5°C; in this case, the value is -0.53°C which is insignificant. The cokriging method might not have the largest region of accurate data but it would still be considered as the better model of the three because it is accurate in the region that matters. Greater Vancouver area is heavily populated and which is why accurate weather interpolation should be present.

To prove cokriging is a better model than the other two methods; the error reported from this model needs to be inspected. In fact, the cokriging model error is lower than the other two. Root mean squared standardized error shows if the model is under or overestimating the variability in the predictions and optimally should be around 1. A root mean squared standardized error of 0.9 is given for the cokriging model whereas 1.58 is given for the kriging model and 2.02 is given to the IDW model. This root mean squared standardized error shows cokriging is slightly underestimating the variability whereas the other two methods are substantially overestimating the variability. The mean standardized error shows the average of the standard deviation of the distribution and optimally should be around 0. A mean standardized error of -0.6 is given to cokriging model and -0.12 is given to the kriging model (not available in IDW model). That shows the kriging model has twice the standardized error compared to the cokriging model. For the above reasons, the cokriging method in this project is preferred over the other two in this project.