Background points: are they mandatory?
Trevon Fuller
due to the species' low abundances. Modeling the fundamental niche of
such species requires the species' occurrence points. In some cases,
it may also be appropriate to generate pseudo-absences for the
species. However, since pseudo-absences can be criticized as
arbitrary, ideally, the model of the fundamental niche should be
constructed using a method that only requires the species' occurrence
points.
Among the stated advantages of the Maxent approach is that "[i]t
requires only presence data, together with environmental information
for the whole study area" (Phillips et al., Ecological Modelling 190,
p. 234). Dudík et al.'s chapter in the book edited by Shawe-Taylor
and Singer provides a similar description of the input data required
for the maximum entropy approach. However, Phillips et al. also
explain that the calculation of the AUC requires negative instances or
pseudo-absences, which they term "background points" (p. 244).
The Ecological Modelling article and the book chapter could be
interpreted to mean that background points are not mandatory for the
maximum entropy approach. My question is whether background points are
required for the Java implementation of the maxent algorithm described
by Dudík et al. When using Maxent 3.2.1, I have received an error
message that suggests that background points are mandatory.
If I set "Max number of background points" to 0 in the "Maximum
Entropy parameters" dialog box, then during the Maxent run I receive
the error "No points with data in all layers." The error occurs when
the progress bar reaches 99% when the "Maximum Entropy Species
Distribution Modelling" dialogue displays the message "Extracting
random background and sample data". When I set the "Max number of
background points" to 1, the Maxent run finishes without generating
any error messages. The data that I am using are samples file and
environmental layers from the Maxent tutorial.
William Dickey
Veronica
Hi William,
The short answer to this question is YES, you do need background points for modeling in Maxent.
I believe that the misunderstanding lies in this person’s interpretation of what Maxent can do as a presence-only SDM. Maxent is not a presence-pseudo-absence model. It is a presence-only model that compares and contrasts against the surrounding, non-presence areas and treats them as background points. Again, background points are NOT absences or pseudo-absences. This is because unlike other presence-absence or presence-pseudo-absence SDMs, Maxent does not treat presences as “preferred” areas and the background as “non-preferred” areas. As Franklin stated in her book, ‘Mapping Species Distributions: Spatial Inference and Prediction,’ Maxent “agrees with everything that is known but avoids assuming anything that is not known.” (Chapter 7).
See also Merow et al 2013, ‘practical guide to MaxEnt for modeling species’ distributions: What it does, and why inputs and settings matter’, supplementary material:
“Pseudoabsence: An ambiguous term used to refer to locations where it is unknown whether the species is present or absent. Some authors equate pseudoabsences with MaxEnt’s background sample, while others use modeling strategies to choose locations that are expected to be unlikely to contain the species. We suggest avoiding this term.”
And in the Hijmans and Elith 2013 version of ‘Species Distribution Modeling in R’, they say:
“Background data (e.g. Phillips et al. 2009) are not attempting to guess at absence locations, but rather to characterize environments in the study region. In this sense, background is the same, irrespective of where the species has been found. Background data establishes the environmental domain of the study, whilst presence data should establish under which conditions a species is more likely to be present than on average. A closely related but different concept, that of ”pseudo-absences”, is also used for generating the non-presence class for logistic models. In this case, researchers sometimes try to guess where absences might occur – they may sample the whole region except at presence locations, or they might sample at places unlikely to be suitable for the species. We prefer the background concept because it requires fewer assumptions and has some coherent statistical methods for dealing with the ”overlap” between presence and background points (e.g. Ward et al. 2009; Phillips and Elith, 2011).”
To get even more complicated, this means that background points can technically overlap presence points (!!), but pseudo-absence points cannot.
So where do we go from here?
Well, keep in mind that Maxent’s default is to have 10,000 background points. But, if your study area is small, or the resolution of your dataset causes you to have less than 10,000 pixels available for training, then try lowering that number and seeing how it affects your results. If you worry that potential biases exist with your presence data (e.g., they all occur alongside the roads due to access), create a bias file. It is also important to consider training your model to a small area (the technically reachable areas for the species) and then projecting the prediction to the entire study area (see Elith et al. 2010, ‘The art of modelling range-shifting species’ and their supplementary material). And you can also have your training and testing data not be a subset of the same entire dataset, but from 2 independent datasets (i.e., 75% train from one set of surveys, and 25% train from another set of surveys). There are many options. You just have to look at the literature and see what works for your situation.
I hope this helps, or at least gets the thoughts going on this topic.
Cheers,
Veronica Frans
Center for Systems Integration and Sustainability
Michigan State University
--
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 post to this group, send email to max...@googlegroups.com.
Visit this group at https://groups.google.com/group/maxent.
For more options, visit https://groups.google.com/d/optout.
Felipe Coutinho Maciel
To unsubscribe from this group and stop receiving emails from it, send an email to maxent+unsubscribe@googlegroups.com.
To post to this group, send email to max...@googlegroups.com.
Visit this group at https://groups.google.com/group/maxent.
For more options, visit https://groups.google.com/d/optout.
--
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+unsubscribe@googlegroups.com.