Does Maxent probability change according to layers resolution?

[alaaeldeen80]

Dear Maxent users,

I have used maxent to model the potential disribution for my species,
I have used two spatial resolutions (30sec. and 2.5 min.), same
variables set, same presence points, and same setting were adopted
with each resolution. The distribution range of 2.5min. is larger
than that predicted with 30sec. Does anyone here can give explanation
for that?
Cheers
Alaa Eldin

[Husam El Alqamy]

Dear Alaa
I think this is because Maxent treat the whole cell for the value of the
environmental variable raster to its center. In other words if the presences
record falls in a cell of the 2.5min this will extend a positive response to
this environmental value to 2.5min and accordingly to cells of similar
values for 2.5min each while in the case of 30sec the process will happen on
the extent of 30sec for each cell. Thus resulting in a bigger range using
the same everything in the case of 2.5min compared to 30sec. Therefore finer
resolution is always better for prediction but you have to compromise
according to your presence dataset, computing capacity, etc.
Hope this was helpful.
Regards

Husam El Alqamy, B.Sc, M.Phil.
Sr. Biodiversity GIS Analyst,
Enivronmental Information Sector
Environment Agency - Abu Dhabi

Dear Maxent users,

--
You received this message because you are subscribed to the Google Groups
"Maxent" group.
To post to this group, send email to max...@googlegroups.com.
To unsubscribe from this group, send email to
maxent+un...@googlegroups.com.
For more options, visit this group at
http://groups.google.com/group/maxent?hl=en.

[alaaeldeen80]

Thanks Hussam,
It makes sense for me.
Cheers

[Tropica]

Dear Husam,

I think that without further context it is very dangerous to say that "finer resolution is always better for prediction"

I've said it before in response to another piece on here, but it is worth thinking very carefully about "perceived accuracy" in the finer resolution gridded datasets available and whether or not any apparent "improvements" are ecologically relevant or meaningful.

There are many dangers with equating increased resolution with increased realism!  Quite a few publications question the utility of data sets at a resolution finer than 30' because of the quality of and inherent uncertainty in the raw station data supporting these gridded data products.  Of course we don't know what variables Alaa is using, what the distribution data looks like, how the model has been constructed or where it is being projected to.  Even so, I would be surprised if everything pointed towards 30 seconds resolution being the appropriate way forward.

I suggest reading the following publications for more on this issue:

http://dx.doi.org/10.1175/1520-0442(1999)012<0829:RTCSTC>2.0.CO;2

http://dx.doi.org/10.1002/joc.1322

http://dx.doi.org/10.1111/j.2041-210X.2011.00134.x

Happy modelling!

[Ahmed El-Gabbas]

Dear Tropica, Hussam, and Alaa

Which grain size (30 arc seconds or 2.5 minutes) is more relevant to use when modelling in a relatively large study area (Egypt) using most of the species records extracted from museums and published articles; i.e. with relatively high uncertainty? The environmental variables to be used in the model are the BIO layers downloaded from www.worldclim.org/, altitude, landuse, and NDVI. I would project the predictions to the future to show the effect of climate change on the species and also would prioritize the important areas for my target group ..

If I am going to use 30 arc seconds, what is the maximum uncertainty I should avoid in my data?

Regards,

Ahmed

[Husam El Alqamy]

Dear Tropicaonline and List

 

I agree that there is a considerable deal of uncertainity in the worldclim data especially for Egypt as there are less stattions for the data to originate from. However I am more familiar with Alaa data and modelling approach, thats why I responded as such. The data sets used involved many locally generated environmental variables such as altitude, NDVI, and distances from roads, and these are interpolated on the scale of ~1.0Km so my advice to go for the finer scale as cell influence will be limited to smaller span provided some small sample size of presence records for some of the species under investigation.

Many Thanks for the literature links and hope to hear from you about it.

 

Regards

Husam

--
You received this message because you are subscribed to the Google Groups "Maxent" group.

To view this discussion on the web visit https://groups.google.com/d/msg/maxent/-/ezv4KZii-EcJ.

To post to this group, send email to max...@googlegroups.com.
To unsubscribe from this group, send email to maxent+un...@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/maxent?hl=en.

--

Husam El Alqamy, B.Sc., M.Phil.
Sr. Biodiversity GIS Analyst ,
Environmental Information Sector, EIS
Environmental Agency Abu Dhabi,UAE
Antelope Specialist Group, ASG - IUCN

[alaaeldeen80]

Dear Husam and Tropicaonline,

Thanks for your interest and response.
Your points are worth thinking, but it bring up another point which
highlighted in Ahmed's post, which scale is recommended with
occurrence data with relatively high uncertainty level such as the
data collected from museums and publications?
Thanks in advance.

Alaa

> >http://dx.doi.org/10.1002/joc.1322<http://www.google.com/url?sa=D&q=http://dx.doi.org/10.1002/joc.1322&u...>
> >http://dx.doi.org/10.1111/j.2041-210X.2011.00134.x<http://www.google.com/url?sa=D&q=http://dx.doi.org/10.1111/j.2041-210...>

[Tropica]

Hi,

Ahmed, I think the question should perhaps not be what is the maximum uncertainty that should be avoided in the [distribution] data, but rather what is the likely impact of quality variation in the distribution data on model output given the other sources of uncertainty present in the model.  There are obviously trade-offs involved.  I also wonder if using land use as a covariate for projecting into the future is useful.  Do you have a way for factoring in future representations for land use (it doesn’t strike me as an obvious choice for a static variable)?

As Husam points out, there are sources of data other that WorldClim that may be better to use for your modelling.  At a global level, the CliMond dataset has improved on some of the limitations with the WorldClim data and at a regional level there may be other datasets available with additional variables that are more ecologically relevant to your modelled species and the scale of the model you are constructing.

The problem with the term relative (as in “relatively high uncertainty”) is that, well, it’s relative!  This reinforces my initial point that everyone will need to consider their own particular situation and the unique context it will inevitably have.  There is certainly no one resolution rule to fit all.  I see that Bruce Miller makes a similar point on this thread: http://groups.google.com/group/maxent/browse_thread/thread/0c8918c3b0075ac8#

As, for example, Elith & Leathwick (2009; http://dx.doi.org/10.1146/annurev.ecolsys.110308.120159) point out, choice of resolution depends on issues such as the spatial accuracy of the data, characteristics of the terrain, life history and putative range determining factors for the species, and the intended application of the modelling exercise.

Do keep in mind that (i) many of the GCM anomaly surfaces are almost 2° resolution in their native form, (ii) in many parts of the world the station data underpinning the global gridded datasets are hundreds of kilometres apart (e.g. Fig 1 in http://dx.doi.org/10.1002/joc.1276), and (iii) that topographic heterogeneity can lead to steep covariate gradients over short geographical distances.

I’ll say again that when modelling over large areas, what appears to be a default choice in many papers of using the finest available gridded data without careful consideration seems a surprising one.  Perhaps the one exception I can think of where dropping down to finer scales makes sense is where high topographic variation may be influencing the projection suitability at its range margins (i.e. a finer resolution would allow the user to pick up on this variability).

Either way, I would recommend reading the section entitled “Spatial Scale” in Elith & Leathwick (2009) and some of the references in their Supplemental Literature as a starting point for a better understanding of these issues.

I hope that helps a little bit and good luck!