Density Surface Modelling: grid cell size vs segment size

256 views
Skip to first unread message

Adrian Schiavini

unread,
Aug 8, 2014, 3:40:59 PM8/8/14
to distance...@googlegroups.com
Hi everybody,

I wonder if there is a rule of thumb regarding the relationship between the size of the segment (when you split the transect in segments) and the cell size for the prediction grid.

There is a recommendation that the segment should have twice the ESW or the strip band width in length.

On the other hand, environmental covariables may come from different sources with different resolutions: for exampe, a Digital Elevation Model comes with a 90m resolution and NDVI data from MODIS comes in 250m resolution. Even when we can average this values under a segment of a larger size, one would expect that using a smaller grid cell size would produce a more precise abundance estimate, but surely there is a trade-off.

Any suggestions welcome...

Eric Rexstad

unread,
Aug 10, 2014, 10:00:51 AM8/10/14
to Adrian Schiavini, distance...@googlegroups.com
Good questions Adrian.

We do not have 'conventional wisdom' on this subject that differs from yours.  You have touched on the important topics.  As you note, often the covariate data you wish to use are from other studies, unrelated to your organism.  This can result in a heterogeneous assemblage of data, none of which was necessarily collected for the use you will put it to.  You would like to use a spatial resolution that is relevant to your study organism.

Anxious to hear what arises from this discussion.
--
You received this message because you are subscribed to the Google Groups "distance-sampling" group.
To unsubscribe from this group and stop receiving emails from it, send an email to distance-sampl...@googlegroups.com.
To post to this group, send email to distance...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/distance-sampling/a29c28ea-437d-4667-b02b-3d48e2d2c579%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

-- 
Eric Rexstad
Research Unit for Wildlife Population Assessment
Centre for Research into Ecological and Environmental Modelling
University of St. Andrews
St. Andrews Scotland KY16 9LZ
+44 (0)1334 461833
The University of St Andrews is a charity registered in Scotland : No SC013532

Sarah Moore

unread,
Dec 28, 2015, 2:45:19 PM12/28/15
to distance-sampling, aschi...@wcs.org, eric.r...@st-andrews.ac.uk, er...@st-andrews.ac.uk
Hello,

I am currently struggling with this problem. My ESW is a few meters. My grid cells are 30m. I have produced a DSM but my model is terrible. I really think that my segments are too small (4 meters or so) which means I have a ton of segments (each transect is 2km and I have close to 200 transects.  This results in several segment centroids being on the same pixel and thus having the same value leading to data clumping. My next step is to split the transects into 90m segments which would give at least a pixel between each centroid. Is this a bad idea? Any suggestions?

Sarah

Adrian Schiavini

unread,
Dec 29, 2015, 8:04:19 AM12/29/15
to distance-sampling
Hi Sarah,

What do you mean with "terrible model"? Poor deviance explained? Crazy extrapolation in numbers?

Besides the issue you mention, that may lead to some kind of pseudo replication of segments and their subsequent covariates, I wonder if your transects sampled the range of the covariates used for the model. 

Sometimes if you sample part of the range and you find a good relation, when you extrapolate to the whole range you may obtain wierd extrapolations, together with large CVs.

It will also help to understand that you let us know the objet of the study and the area where you want to estimate.




Click here to report this email as spam.




--
Dr. Adrián Schiavini
Ecología y Conservación de Vida Silvestre
Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET)
Centro Austral de Investigaciones Científicas (CADIC)
Bernardo Houssay 200
(V9410BFD) Ushuaia
Tierra del Fuego
Argentina
 
Teléfono / Phone: +54 (2901) 422310/314, 433320  ext. 131
FAX: +54 (2901) 430644.
E-mail:   aschi...@wcs.org
             aschi...@gmail.com
             aschi...@cadic-conicet.gob.ar
Web: www.cadic-conicet.gob.ar

Eric Rexstad

unread,
Jan 4, 2016, 3:21:58 AM1/4/16
to distance-sampling
Thanks for your comments Adrian.

Going back to Sarah's question about segment size relative to covariate resolution (I take it that is what 'pixel' was referencing).

This is a philosophical problem, but one that deserves thought.  The concept is to model the animals' relationship with their environment.  But easier said than done.  Sampling of the animals is being done via the transect survey whereas sampling of the environment is being done by different researchers (from outer space).  In a perfect world, those two investigations would be conducted at the same scale, but this is life as we know it.

Sarah's suggestion of re-segmenting such that the animal sampling is 'scaled up' to the more coarse environment sampling is probably a sound choice.

For more options, visit https://groups.google.com/d/optout.

Rob Williams

unread,
Jan 4, 2016, 12:23:51 PM1/4/16
to Eric Rexstad, distance-sampling
Hi,
This post seems to conflate a number of issues, and I’m not clear which is the biggest problem.  If the issue is over dispersion (e.g., if the vast majority of short segments contain zero sightings), it’d be worth trying a Tweedie distribution in addition to trying longer segments.  That helped in two of our studies that had a lot of zeroes:
http://www.nature.com/articles/srep04170

If there are many sightings, most segments contain sightings, and the poor fit is due to variability in environmental conditions along long segments, or lack of variability in environmental conditions among short segments, that’s a different problem with a different solution.  If it’s a tiny study area, maybe environmental conditions didn’t vary much throughout the survey region, and you need more informative covariates.  Maybe that’s poor model fit is the answer — variability in X across the range of values of X you observed, doesn’t explain much variability in encounter rate.

Choice of segment length was raised as a priority area for future research in Hedley’s original thesis: http://www.creem.st-and.ac.uk/sharon/public_html/thesis.pdf  Hedley recommended checking how sensitive your model fit is to choice of segment length. If segment length is causing your answers to vary by a lot, is it worth choosing one of Hedley’s other methods, such as the waiting distances method, which is less reliant than the count method on choice of segment length?  Adrian raises a good point:  density surface models relax a lot of assumptions about sampling design, but they still assume reasonable coverage across the range of all covariates in the model.  If that assumption was violated, I could imagine poor fit and wild extrapolations.  

I think a bit more detail is needed to narrow down the laundry list of potential explanations, so people can offer more helpful advice.  

All the best,
Rob




Sarah Moore

unread,
Jan 10, 2016, 4:32:51 PM1/10/16
to distance-sampling, Eric.R...@st-andrews.ac.uk, rm...@st-andrews.ac.uk
Hello All,

Originally my segments were around 4 m each (due to small ESW). This was leading to around 50,000 segments and a lot of segment centroids on the same pixel (30m spatial resolution for covariates). I tested a segment size of 30 m and a segment size of 90 m and found that the models, negative binomial and tweedie both performed better (higher AIC and more deviance explained) than the original 4m or 30m segments. I also tested dividing up my study area and found that by focusing on either half of the study area, I increased deviance explained by 50%. So, Adrian is correct in that maybe my segments (when modeled over the entire study area) are not covering the full range of the covariates used in the model. However, I do recommend paying attention to the spatial scale of covariates when determining segment length!  I went from around 8% deviance explained to 20-25% deviance explained by increasing segment size.

Peace, Sarah
Hi,
--
You received this message because you are subscribed to the Google Groups "distance-sampling" group.
To unsubscribe from this group and stop receiving emails from it, send an email to distance-sampl...@googlegroups.com.
To post to this group, send email to distance...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/distance-sampling/CAAtCR4zqYdPnueHP5t4%2BHkyy-4wxxf9_suHiTro2nKQdYuwF9w%40mail.gmail.com.
For more options, visit https://groups.google.com/d/optout.

-- 
Eric Rexstad
Research Unit for Wildlife Population Assessment
Centre for Research into Ecological and Environmental Modelling
University of St. Andrews
St. Andrews Scotland KY16 9LZ
+44 (0)1334 461833
The University of St Andrews is a charity registered in Scotland : No SC013532
Reply all
Reply to author
Forward
0 new messages