Re: [ai-geostats] variograms
If your data is irregularly spaced, then you need to
experiment with 'lag' intervals to balance between
(a) getting enough points to see the shape of the
(b) getting enough pairs in each point to have some
confidence in it.
Remember that each point on your graph is an estimate
of a variance. Some books give hard-and-fast rules
like "you have to have 25 pairs in each point" but,
personally, I think this is fatuous. The real
situation is a bit circular -- if you have a regular
phenomenon, you can get the shape with few samples and
few points; if you have an erratic phenomenon you need
many samples and lots of points.
Over the years, I have found the folowing useful:
i) look at the 'nearest neighbour' or inter-sample
distances to see what the 'natural' spacing in your
ii) Use that to guide your first choice for lag
interval and experiment around that distance.
iii) Use the Cressie goodness of fit statistic to help
you judge the fit of your model.
iv) Use cross validation to help you judge the fit of
the model and the behaviour of the kriging errors.
If your data is on a grid, life is a lot easier, just
use 1/5th of the grid spacing as your lag interval.
The usual rule of thumb on number of lags is not to go
more than half the extent of your study area. That is,
if your study area is 1km on a side, construct your
semi-variogram to a maximum of 500 metres.
Hope this helps
___________________________________________________________ALL-NEW Yahoo! Messenger - sooooo many all-new ways to express yourself http://uk.messenger.yahoo.com