[R-sig-Geo] [Soc] GRASS module for kriging - call for users [was: R-sig-Geo Digest, Vol 69, Issue 21]
Edzer Pebesma
edzer.pebesma at uni-muenster.de
Mon May 25 15:21:15 CEST 2009
Ebrahim Jahanshiri wrote:
> for kriging a preliminary analysis is important. We should know if the data
> has trend or how are the ouliers? should we remove them or not. Some of the
> kriging techniques like universal kriging create very odd values at some
> points if we dont deal with the data frist hand before kriging. I hope that
> we can come up with the software that does all this by itselft ( some sort
> of decision maker like AI). I hope to see this kind of sotware pretty soon
> and I wish it be open source.
>
I agree that this would be nice.
We discussed this issue quite lengthy in the intamap consortium (see
http://www.intamap.org/ , and the wiki as well -- also interesting for
the SoC project). We concluded that for a generic automatic
interpolation algorithm users don't want this to be done automatically,
because no automatic algorithm can decide whether an outlier indicates a
measurement error or a disaster. Remember the ozone layer was overlooked
for many years because an automatic "outlier" algorithm was built in the
data processing chain.
Only if your application is domain specific you can do something useful,
e.g. use physical limits -- soil moisture content can never exceed 100%;
content variables need to be non-negative.
Of course you can add options, and build optional outlier filters. Then,
how should outliers be marked? Is there any agreement on this? What
would be the defaults for necessary parameters?
Automatic trend detection is another issue that is very interesting.
> On Mon, May 25, 2009 at 6:42 PM, Edzer Pebesma <
> edzer.pebesma at uni-muenster.de> wrote:
>
>
>> I believe that many interpolation problems "out there" are simple, and
>> can be solved using geostatistics with a "finish" or "interpolate"
>> button. A question I find interesting is whether this button should
>> always do its best, or should it be so clever to warn the user in case a
>> problem is not "that simple", and what the criteria are for this.
>>
>> Of course this SoC project is about an "interpolate" button using open
>> source software, so the user is always able to find out every detail
>> behind it. But will she or he actually do this in a non-simple case?
>>
>> As of your claim that no two geostatistical packages provide the same
>> result, as a more optimistic person I'd like to provide some
>> counter-evidence; please run:
>>
>> require(gstat)
>> require(geoR)
>> xyz = data.frame(x = c(0,0,1), y = c(0, 1, 1), z = c(1,2,3))
>> coordinates(xyz)=~x+y
>> x0 = SpatialPoints(data.frame(x=0,y=.5))
>> kr1 = krige(z~1,xyz,x0,vgm(1, "Exp", 1))
>> kr2 = krige.conv(as.geodata(xyz), locations=coordinates(x0),
>> krige=list(cov.model="exponential", cov.par=c(1,1)))
>> kr1
>> c(kr2$predict, kr2$krige.var)
>> kr1[[1]] - kr2$predict
>> kr1[[2]] - kr2$krige.var
>>
>> I'd be more than happy if someone could repeat this, or another example,
>> with ArcGIS or other software, of course.
>> --
>> Edzer
>>
>> Ebrahim Jahanshiri wrote:
>>
>>> Dear Ann,
>>> I think having a "finish" button is the ulitimate thing that all of us
>>> looking for in a software!. Unfourtunately it is not possible. There are
>>>
>> far
>>
>>> too many issues regarding a good kriging operation that prevents you from
>>> trusting a nice package like "Geostatistical Analyst" in ArcGIS. The
>>> evidence is that you cant find two geostatistical package that their
>>>
>> resutl
>>
>>> is the same (I know this is a big statement but I have seen this for a
>>> couple of softwares!). I cant say the surface that is generated by ArcGIS
>>>
>> is
>>
>>> not accurate but it bothers me to not know what happens beneath the
>>>
>> program
>>
>>> when it fits the variogram and calculates the matrices for kriging and
>>>
>> more
>>
>>> importantly a "routine" that you mentioned will not work for all the
>>>
>> data.
>>
>>> So personaly I will go for the chors of GeoR or Gstat instead of using
>>>
>> the
>>
>>> ArcGIS. Though it is not complete and I have to admit that I have a lot
>>>
>> of
>>
>>> problem adapting to it.
>>> Ebrahim
>>>
>>> On Sun, May 24, 2009 at 1:59 AM, Anne Ghisla Insubriae <
>>> a.ghisla at studenti.uninsubria.it> wrote:
>>>
>>> Ebrahim Jahanshiri a écrit:
>>>
>>>>>> Dear Anne,
>>>>>>
>>> Ebrahim,
>>>
>>>
>>>>>> Thanks for the enquiry. I think we should have done this long before.
>>>>>>
>> I
>>
>>>>>> havent done any geostatistics with GRASS but I did alot in ArcGIS and
>>>>>>
>>> also
>>>
>>>>>> in R (variography). I have to tell you that ArcGIS is for those who
>>>>>>
>> want
>>
>>> to
>>>
>>>>>> have a surface only by just clicking on "finish" button trought its
>>>>>> powerfull interface. That is, it does not provide a good scientific
>>>>>> backgroud for you to check the procedure. for example how it fitted
>>>>>>
>> the
>>
>>>>>> variogram model or how exactly it does the "validation". That is why
>>>>>>
>> the
>>
>>>>>> "Geostatistical package" objects or classes are not available for the
>>>>>> programmers to program with (there is a kriging in "Spatial analyst"
>>>>>>
>>> package
>>>
>>>>>> though that its object is programmable and I personally did some with
>>>>>>
>> it.
>>
>>>>>> but it lacks the variogram modeling). I think they are still working
>>>>>>
>> on
>>
>>> it.
>>>
>>>>>> In short I dont recommend ArcGIS for the so called scientific kriging.
>>>>>>
>>> I was afraid of this. Therefore I guess an interface with too many
>>> mandatory options would be considered overcomplicated.
>>> There should be as more options as needed for a flexible calculations,
>>> and also proper defaults to clik on Ok and get an acceptable result.
>>> What do you think about this solution?
>>>
>>>
>>>>>> In
>>>>>> the other hand R is very powerfull in that it gives you freedom to do
>>>>>>
>>> your
>>>
>>>>>> own style through coding. I worked with "GeoR" pakcage and it is
>>>>>>
>> pretty
>>
>>> nice
>>>
>>>>>> both in terms of variography and kriging. There are other package like
>>>>>> "Gstat" that I havent worked with but have good qualities. GeoR
>>>>>>
>>> implements
>>>
>>>>>> the so called "model based geostatistics" which is the application of
>>>>>> bayesian statististics to the geostistics I guess and it is quite new.
>>>>>>
>>> I'm presently having a look at both geoR and gstat, and also the wrapper
>>> package automap.
>>> I guess that the first users of the new module would be already familiar
>>> with R and GRASS, so their feedback about most used functions is very
>>> important.
>>>
>>>
>>>>>> I hope these comments were helpful and please let me know if you need
>>>>>>
>> any
>>
>>>>>> further explanation from me,
>>>>>>
>>> Very helpful, thank you very much!
>>>
>>>
>>>>>> Ebrahim
>>>>>>
>>>>>>
>>> best regards,
>>> Anne
>>>
>>>
>>> ------------------------------------------------------------------------
>>>
>>> _______________________________________________
>>> R-sig-Geo mailing list
>>> R-sig-Geo at stat.math.ethz.ch
>>> https://stat.ethz.ch/mailman/listinfo/r-sig-geo
>>>
>> --
>> Edzer Pebesma
>> Institute for Geoinformatics (ifgi), University of Münster
>> Weseler Straße 253, 48151 Münster, Germany. Phone: +49 251
>> 8333081, Fax: +49 251 8339763 http://ifgi.uni-muenster.de/
>> http://www.springer.com/978-0-387-78170-9 e.pebesma at wwu.de
>>
>>
>
>
>
>
--
Edzer Pebesma
Institute for Geoinformatics (ifgi), University of Münster
Weseler Straße 253, 48151 Münster, Germany. Phone: +49 251
8333081, Fax: +49 251 8339763 http://ifgi.uni-muenster.de/
http://www.springer.com/978-0-387-78170-9 e.pebesma at wwu.de
More information about the R-sig-Geo
mailing list