[R-sig-Geo] CODE for spatial logistic regression

[Alexander Brenning]

Hi,

landcover classification with four classes... though I haven't 
experimented with spatial logistic regression, I would expect that there 
is more information in the spectral satellite data (feature space) that 
you are probably using as predictors than in the adjacency/distance 
information. The very nature of a categorical response variable implies 
that the response is discontinuous in geographic space, so a 
kriging-type approach would only help if you had high sampling density 
in your learning sample (several samples within each homogeneous 
landcover polygon - not a typical situation).

The feature-space perspective - ignoring geographic space - allows you 
to look at many excellent classification techniques that make best use 
of your predictor variables; have a look at the book of D.J. Hand: 
"Construction and assessment of classification techniques" for an 
overview. On the other hand, spatial continuity is often achieved by 
applying filters to the resulting prediction image.

Here a recent comparison of different classification techniques in crop 
identification:
http://www.zfl.uni-bonn.de/earsel/papers/64-71_brenning.pdf

Multinomial (non-spatial) logistic regression was pretty, support vector 
machine performed better. A comparison with spatial log.reg. would of 
course be interesting.

Maybe the more important spatial issue in using classification 
techniques is the question of error assessment in a spatial context, 
e.g. using a spatial cross-validation, or in the above paper 
cross-validation at the field level.

By the way, I am organizing a special session on Spatial Classification 
at the meeting of the International Federation of Classification 
Societies (IFCS) in March 2009 in Dresden, Germany. Theoretical and 
applied contributions are welcome, abstract deadline November 3rd. See 
http://www.ifcs2009.de/

Cheers
  Alex



ivan valencia wrote:
> Hi HENK
> 
> My data comes from land use classification, a grid with 1km2 resolution.
> I have a bynary classification with 4 types of land use, for each land use type
> have to run a logistic regression with available covariables also at
> the same raster format.
> I want to consider spatial logistic regression...is it possible?
> {}
> ivan
> 
> 2008/10/9 Henk Sierdsema <Henk.Sierdsema at sovon.nl>:
>> Hi Ivan,
>>
>> Can you tell me what the purpose is of your modelling? Is it simply producing spatial predictions based on a logistic model or do you want to incorporate spatial autocorrelation in the models? Given your last mail it seems you want to incorporate spatial autocorrelation despite the fact that you deny this in your second mail. So please extend more on the type of data you have and your aim. Next to geoRglm, which is only suitable for small datasets, you might also try regression-kriging.
>>
>> Is there by the way anyone who has experience with autoregressive models?
>>
>> Henk
>>
>>
>>
>> Henk Sierdsema
>>
>> SOVON Vogelonderzoek Nederland / SOVON Dutch Centre for Field Ornithology
>>
>> Rijksstraatweg 178
>> 6573 DG  Beek-Ubbergen
>> The Netherlands
>> tel: +31 (0)24 6848145
>> fax: +31 (0)24 6848122
>>
>>
>>
>> -----Oorspronkelijk bericht-----
>> Van: ivan valencia [mailto:liov2067 at gmail.com]
>> Verzonden: woensdag 8 oktober 2008 17:16
>> Aan: r-sig-geo at stat.math.ethz.ch
>> Onderwerp: [R-sig-Geo] CODE for spatial logistic regression
>>
>>
>> Hello
>>
>> I need some guide about spatial logistic regression, Is it available a
>> code in R?
>>
>> {}
>>
>> LIOV
>>
>> _______________________________________________
>> R-sig-Geo mailing list
>> R-sig-Geo at stat.math.ethz.ch
>> https://stat.ethz.ch/mailman/listinfo/r-sig-geo
>>
> 
> 
> 

-- 
Alexander Brenning
brenning at uwaterloo.ca - T +1-519-888-4567 ext 35783
Department of Geography and Environmental Management
University of Waterloo
200 University Ave. W - Waterloo, ON - Canada N2L 3G1
http://www.fes.uwaterloo.ca/geography/faculty/brenning/