Are special interpolation techniques required for weather data?

Question

I've downloaded a series of weather observations from the local weather agency - it comprises around 1,000 points spread non-uniformly across Australia, with a greater density of points located near populated areas.

I'm interested in displaying mean temperatures and monthly rainfall (separately) as smooth surfaces, so I used the Kriging function in Spatial Analyst (with the default values) to calculate this grid:

[edit]

The grids will be used for visualisation purposes only. In a web map I'll display the grids to give context, but will only allow identify on the actual weather stations, in which case I'll show the original values. I guess this means the accuracy of the grids isn't of paramount importance.

My questions:

• Which of the parameters on the Kriging tool do I need to understand, in order to create meaningful interpolations of rainfall, temperature and other weather factors?
• Do I need to account for effects of topography (eg, orographic rain effect) or does the kriging algorithm handle this?
• is the Geostatistical Analyst a better tool for this type of analysis? (I believe it's designed to handle biases due to geography)
• anything else I should be aware of to create a meaningful result?

Thanks

Answer 1

• You don't tell us anything about the usage of these maps. Do you just want to produce a map? Will you use them as a model input? This may change your methodology.
• Which error is acceptable for the usage you need?
• Topography can only be included if you use it as an input variable, in a procedure called regression kriging. I don't know whether arcgis supports this, but arcgis can not take it into account if you don't supply the information. In this specific case it seems you a lot of data points in mountainous regions so the actual difference may be small.
• You have sufficient points, so I suppose that arcgis will be able to fit a reasonable variogram. In that case using the automatic parameters is ok (at least better than using eg inverse distance weighting). Of course you should be aware that in regions with little observations the prediction has a large error margin.