Skip to content

kriging as fish swim, not as crows fly

10 messages · Martin Renner, Michael Sumner, Pilar Tugores Ferra +3 more

#
Hi All,

I want to kirg fish and seabird densities within an estuary which has several arms. Since neither organisms cross land, the appropriate distances would not be euclidian but over-water (as fish swim). There are several papers, describing this problem and how to deal with it (see below), but I have not found an easily accessible implementation. Is anybody aware of a solution in R?

Best,
Martin



@article{Rathbun:1998aa,
	Author = {Rathbun, Stephen L.},
	Journal = {Environmetrics},
	Number = {2},
	Pages = {109--129},
	Title = {Spatial modelling in irregularly shaped regions: kriging estuaries},
	Volume = {9},
	Year = {1998}}

@article{Little:1997aa,
	Author = {Little, Laurie S. and Edwards, Don and Porter, Dwayne E.},
	Journal = {Journal of Experimental Marine Biology and Ecology},
	Number = {1},
	Pages = {1--11},
	Title = {Kriging in estuaries: as the crow flies, or as the fish swims?},
	Volume = {213},
	Year = {1997}}




Martin Renner				
US Geological Survey			
Alaska Science Center
#
Not kriging as such, but check out the soap-film smoothing in package mgcv:

http://www.maths.bath.ac.uk/~sw283/simon/papers/soap.pdf

FWIW, there are binning methods with MCMC in the package
tripEstimation that have similar features, but they are particularly
focussed on individual track estimation and probably not easily
applied. Is location uncertainty a big issue for your data? What are
the input locations?

Cheers, Mike.


On Wed, Jan 27, 2010 at 7:06 PM, Martin Renner
<martin.renner at stonebow.otago.ac.nz> wrote:
#
Hi, Martin
Two years ago there was a similar discussion on the list:
http://markmail.org/message/tw7ulpxbgixieyys

If I understand it properly, your problem is you need to compute your variogram using non-euclidian distances.
As far as I know, there is no package in R that can do so (correct me if I am wrong). 
There is a function LCP_Krige that I think you can find on the net. It is not an R function but a mixture of ArcGis 8.x and Matlab. Maybe you can use it or try to adapt it to R...
Jensen, OP, MC Christman and TJ Mller. 2006. Landscape-based geostatistics: a case study of the distribution of blue crab in Chesapeake Bay. Envirometrics. 17:605-621.
I was interested in applying this sort of analysis but I didn't suceed.
I hope you are more lacky!

Cheers,
Pilar 



M? Pilar Tugores Ferr?
PhD Student
Instituto Espa?ol de Oceanograf?a
Centro Oceanogr?fico de Baleares
Muelle de Poniente s/n
07015 Palma de Mallorca
Baleares, Espa?a
Telf.: (34) 971 133759


-----Mensaje original-----
De: r-sig-geo-bounces at stat.math.ethz.ch [mailto:r-sig-geo-bounces at stat.math.ethz.ch] En nombre de Michael Sumner
Enviado el: 27 January 2010 10:29
Para: Martin Renner
CC: r-sig-geo at stat.math.ethz.ch
Asunto: Re: [R-sig-Geo] kriging as fish swim, not as crows fly

Not kriging as such, but check out the soap-film smoothing in package mgcv:

http://www.maths.bath.ac.uk/~sw283/simon/papers/soap.pdf

FWIW, there are binning methods with MCMC in the package
tripEstimation that have similar features, but they are particularly
focussed on individual track estimation and probably not easily
applied. Is location uncertainty a big issue for your data? What are
the input locations?

Cheers, Mike.


On Wed, Jan 27, 2010 at 7:06 PM, Martin Renner
<martin.renner at stonebow.otago.ac.nz> wrote:
_______________________________________________
R-sig-Geo mailing list
R-sig-Geo at stat.math.ethz.ch
https://stat.ethz.ch/mailman/listinfo/r-sig-geo
#
Hi Martin.

I have succeeded in doing something like that.
Please see http://dx.doi.org/10.1016/j.mcm.2009.05.021.

I used GRASS for the computation of the non-euclidean distances and 
modified the geoR library in order to be able to estimate de variograms 
and perform kriging prediction with those "customized" distances.

Both the GRASS script for computing distances and the modified geoR 
library are open source and available at 
http://www.geeitema.org/guenmap/index.jsp?opcion=resultados


I hope it helps.
	?acu.-


Martin Renner escribi?:
#
Hi Facu,

great, this is just what I have been looking for. You say that your modifications to geoR are open source. On your website I found the geoR.dll file, but no source code. The .dll is of little use platforms other than windows. Is there any way I could get the source? 

Thank you for the help! Best,

Martin
On 28 Jan 2010, at 03:42 , Facundo Mu?oz wrote:

            
#
Another approach that might be simpler (or it may oversimplify and not give good results) is to compute the distances between your points as the fish swim, then use multi dimensional scaling (cmdscale function or others) to get a set of points that represent those distances and do the rest of your analysis on the transformed points.
#
Hi Gregory,

An ingenious idea! I guess it would work but I'd be a little bit concerned about the alignment of the predicted surface. For now, I'll try Facu's method and see how that goes.

Best,
Martin
On 29 Jan 2010, at 08:11 , Greg Snow wrote:

            
#
Multidimensional Scaling (MDS) is certainly an option to be considered.
Its suitability depends, of course, on the specific application.

In my case, which is conceptually similar to Martin's, that option was 
discarded.
And I think I can explain why whith this single image:

http://www.geeitema.org/doc/guenmap/docs/distances.gif

Here, three types of distances between all pairs from a set of points 
are compared.
The reference distance (red) is the "cost based" distance (or "water 
distance"), which is the minimum distance measured within the 
"permisible" region (i.e. "as the fish swims").
In black are represented the corresponding euclidean distances, while in 
blue, the corresponding MDS distances (in 2D).

As you can see, the MDS approach corrects the "bias" in the long terrm.
However the variability around the "correct" value holds.

So, while it provides the best overall euclidean approximation, it does 
not account for the behaviour due to the shape of the region.
Since the latter was what we was looking for, we discarded this approach.

One could try to use MDS with higher dimension. But it does not improve 
things very much, specially if you have regions with "holes".
Distances in such regions don't have an euclidean representation 
regardless of the dimension.

Bests
?acu.-


Greg Snow escribi?: