Deciding number of background points

22 views
Skip to first unread message

Ione Arbilla

unread,
Sep 24, 2021, 4:38:27 AMSep 24
to Maxent
Dear All,

I am running maxent models in r with the ENMeval package. 

I have around 14.000 presence points (collected from random sampling using satellite imagery ) and the whole presence area of the species is 34.000square kilometres. The resolution of the raster layers is 90m (the total number of raster cells for each layer is 4.600.000 cells). I think this could be a situation in which the default of 10.000 background points could be changed, but I am having difficulties justifying what value would be appropriate. Is there any method I could use to calculate how many background points would be acceptable? Any ideas are welcome,

Best regards,

Ione

Cooper Marcus

unread,
Sep 24, 2021, 11:17:09 AMSep 24
to max...@googlegroups.com
Have you tried a range of background point settings and examined the resulting model performance scores?

I know that "I picked the background points to maximize model performance" isn't a great justification by itself, but maybe it is part of it?

I've done some experimenting with our model using varying background settings, and tracking run time and AUC scores (each pair of rows below is train then test for various measures of performance - all the numbers below express either test or train AUC of a single model config - the only thing that is changing is background setting)

image.png


--
You received this message because you are subscribed to the Google Groups "Maxent" group.
To unsubscribe from this group and stop receiving emails from it, send an email to maxent+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/maxent/08b29d2a-245b-44d9-910c-e582139b93b5n%40googlegroups.com.

Ione Arbilla

unread,
Sep 26, 2021, 9:59:21 PMSep 26
to max...@googlegroups.com
Hello Cooper,

Thank you so much for your answer! I have run the model with a set of different background points as well, but making a table to compare properly is a great idea. 

 I understand that when it comes to background points there is a moment when it makes no sense to add more (when the performance does not change significantly) but I am wondering l if there is a rule to calculate the "optimal" amount for a certain area extension/occurrence points number and when the change in the performance stops being significant. Anyway, thank you again for your ideas!

Ione

Hau idatzi du Cooper Marcus (cooper...@presencepg.com) erabiltzaileak (2021 ira. 25, lr. (00:47)):

Belay Ejigu

unread,
Sep 27, 2021, 5:09:10 AMSep 27
to max...@googlegroups.com
Reply all
Reply to author
Forward
0 new messages