RE: [ai-geostats] Spatial analysis
- Good day everyone !
"It is true that the general consensus is that you need at least 50 to
100 points for a stable semi-variogram, but myself i bent this rule a
little and i worked with 35 points .... "
.... and I even worked with 9 or 10 points recently (how shocking !).
It is true that one of the favourite comments made by reviewers (those
unexperienced in geostatistics) is "how many points do you need/did you
use" and the favourite reply of the authors is almost systematically
"there is a rule of thumb about this matter... 30 pairs of points is a
minimum" followed by a reference to Cressie's book. Such a reply is a
convenient way to get around the discussion. I must admit, I also used
the same reference to avoid wasting time and space with discussions that
would distract the reader from the main argument treated.
Practically, the number of points you need to describe a spatial
correlation is very much depending on your experience with the
phenomenon you study. If you expect a strong correlation (because of
previous case studies, litterature) and see such a correlation, even if
calculated on very few observations, why not trust your results? If you
don't see anything, than you could consider increasing your sample size.
Few observations will obviously not lead you to results that are
statistically robust but they may be enough to confirm your assumptions.
I believe everyone who has a bit of experience in geostatistics has been
impressed by the large amount of silly semi-variograms published in peer
reviewed papers: even if these semi-variograms were calculated on
thousands of points, many are showing pure noise. Nevertheless, some
dare fitting a model on these semi-variograms and derive conclusions
and/or generate maps.
My 2 cents contribution for the week ...