RE: [ai-geostats] practical range vs range
- Hi Els,
To add to Digby's comments.
For a given block you can run several estimates, each with a different
search ellipsoid dimension. In my case, I tend to run 3 estimates (passes)
for each block being estimated:
1st pass = 0.5 x range in the XYZ dimension
2nd pass = 1.0 x range in the XYZ dimension
3rd pass = 2.0 x range in the XYZ dimension
I have higher confidence in the 1st pass, moderate confidence in the 2nd
pass, and lower confidence in the 3rd pass.
I am sure that many people have several variants of this - there are no hard
rules here, simply what suites your data and its setting. I would however
caution you not to make your sample search ellipsoid dimensions too large. I
have seen a few instances where because of this, samples located near the
very edge of the ellipsoid are assigned negative weights, and the result is
a very small negative block estimate.
From: Digby Millikan [mailto:digbym@...]
Sent: 23 March 2005 11:46
To: ai-geostats; Els Verfaillie
Subject: Re: [ai-geostats] practical range vs range
Which GSLIB program are you using? When you refer to the practical range a
and 3a this is normally associated with the exponential model, the GSLIB
programs appear to be setup to use spherical model parameters, so you should
fit spherical models to your data, in the case of fitting a spherical model
to your variogram (you can fit nested models if you are not happy with a
single structure spherical model) the input to the programs, e.g. in the
case of kt3d, spherical models only have one range i.e. a.
For the search radius most will tell you the range, but it really depends
on how much confidence you are prepared to have for your estimates, you can
extend your search radius further than your range, it's just that those
points estimated which use values greater than the range distance will have
a lower confidence.
You can even have points estimated which only use data at distances greater
than the range, in which case these estimates will have a low confidence. It
depends on how desperate you are to get estimates into data points on how
far you extend the search radius beyond the range. In mining we classify all
estimates with a confidence, either by associating it with the search radius
that was allowed for a data point e.g. a geologist from visual assesment of
the continuity of the geology of an ore zone may draw a polygon extending
10m either side of the drillhole, and say in that case that everything in
that polygon may fall into the Joint Ore Reserves Committes code as the
classification as "Measured which means the entire polygon has the go ahead
for mining depending on its economics, in which case the search radius would
be extended beyond the range, if necessary, so that all blocks within that
polygon get filled with grade. Note also that the confidence of the estimate
at each point is provided by the kriging variance, so if you do extend the
search radius beyond the range, you will have the kriging variance as
another method of classifying the resource.
i.e. You don't have to limit your search radius to the range, it's just
that estimates based on samples using some data greater than the range will
have a lower confidence, indicated by the kriging variance, which in some
cases may be better than having no estimate.
- Hi Els,
The key question here is the sampling density and how many data will
be included in this search window. If there are many, the screening effect
will greatly attenuate the impact of the data further away, hence using a or
3a won't make a big difference. If data are sparser, then usually I set up my
search strategy in terms of maximum number of data, not maximum search
radius, at least in 2D (in 3D setting the search ellipsoid right is very important).
Although simple kriging weights become zero beyond the range, it is not
the case for ordinary kriging, which is a reason why you shouldn't systematically
discard the observations outside the range of autocorrelation, in particular
if the sampling density is low..
From: Els Verfaillie [mailto:els.verfaillie@...]
Sent: Wed 3/23/2005 5:08 AM
Subject: [ai-geostats] practical range vs range
I want to do ordinary kriging with an anisotropic variogram with GSLIB. My
variogram is an exponential model with a practical range of 1800 m in
direction 50 and 880 m in direction 320. I'm not sure whether I have to use
the practical range (which is 3a) or the value a, which is respectively 733
m and 293 m. Furthermore I wonder which maximum search radius I have to
choose: the 3a or the a value?
Els Verfaillie, PhD student
Renard Centre of Marine Geology - Ghent University
B-9000 Gent - Belgium
tel: +32-9-2644573 fax: +32-9-2644967
No virus found in this outgoing message.
Checked by AVG Anti-Virus.
Version: 7.0.308 / Virus Database: 266.8.0 - Release Date: 21/03/2005