Qualitaive Comparison

There are several broad patterns that are observed in a visual comparison of the three models evaluated (refer to Figures 14, 15, 16, 17, 18). One of the most apparent characteristics is that in each of the figures, the surfaces generated from IDW and Ordinary Cokriging appears to be much more similar in general form than does the precipitation surface generated from the Linear Regression. This is due to the fact that both these models include the spatial distribution of stations as a model input whereas the Linear Regression model does not. However by using the spatial distribution of stations as a model factor, the surfaces estimated are therefore strongly influenced by the actual location of stations. Consequently, areas with poor station coverage are poorly resolved in the predicted surface. This is best exemplified in the northeast corner of the study area, where there is almost no detail offered by the first two models. But in this region, and indeed across the breadth of the northern half of the study area, the Linear Regression model generates quite a high level of surface detail. Elevation data, on which the Linear Regression model is solely based, is evenly available across the entire study area. This allows a consistent amount of detail to be preserved in the estimated surface, whereas the extremely uneven distribution of station locations produces fairly irregular consistency in the level of detail afforded by the interpolated surfaces of the first two models.

There are several unusual features of the surfaces produced by the first two models that can be explained by the heterogeneous distribution of stations. Firstly, all the figures show that for both models the surfaces generated tend to increase towards a pronounced peak in precipitation. This point corresponds with the stations of maximum rainfall, rather than being an expression of precipitation generally increasing to the north, as presented by the Linear Regression model. This is likely a product of the data more than an accurate representation of the true nature of the precipitation surface. Secondly, there is a pronounced discontinuity in the north-central region of the study area, which is present in all monthly and annual surfaces predicted by both models. This is certainly incorrect with respect to the actual precipitation pattern and merely a product of the poor distribution of stations. Points along this feature are approximately midway (according to weights) between their nearest neighbours, which are actually quite distant. The rapid rate of change across this feature is a result of the dramatic difference in precipitation between the nearest neighbours to the northwest and those to the southeast. This feature would not exist if there were sample locations located nearby. Thirdly, there are isohyets that delineate peaks in the estimated surface that correspond with the location of individual surfaces. It is generally not realistic that there are true peaks in the precipitation surface corresponding with these locations so much so that they represent artifacts from the relation between the interpolation algorithm and the spatial location of the stations. This is most evident in the results from the IDW model. All these factors emphasize the importance of model validation. In order to determine the ‘best’ model, RMSE values for the estimated surfaces were compared. However, it is critical to emphasize the restriction imposed by only being able to validate the model at the poorly distributed sample locations – it is impossible to validate the models for the regions between sampled locations. It is possible that the ‘best’ model, as determined by this method of evaluation, may not be the model that most accurately represents the true precipitation surface.

All monthly and annual surfaces generated from the IDW model are presented in a precipitation atlas for the GVRD in Modeling Results.