Hello Bruno,

In reality, for GARP background and absence is the roughly the same concept. All points without presences are pseudo-absences.

For a good overview of algorithms I would recommend the book "Mapping species distributions - Spatial inference and prediction" by J. Franklin. 

Hello Bruno

I'm not specialized in such things but I used (one of my student (Roxanne Leberger) was doing it for me) the package BIOMOD2 and it seems that problems of real/pseudo absences is solved using that.

I send you a recent publication where we were using it.

Hoping it helps !!! lol

Cheers from Luc 

Hello,

I'v been working first with biomod and then biomod2. The thing is, biomod always generates pseudo-absences. This is because biomod generates an ensemble of 9 different algorithms, which have different theoretical approaches and data requirements. Some use, presence-only, others presence-absence and presence-background.

And so, to accommodate with the requirements of all these algorithms, biomod asks the user to generate pseudo-absences, which some algorithms will interpret as real absences, other as pseudo-absences and other as background.

Biomod, per se, is not an algorithm, rather it is a combination.

Thank you both for your comments!

Frederico, that's exactly my goal, to generate an ensemble, but respecting the different requirements of each algorithm. I have not yet tried biomod, but I am currently experimenting with dismo.

I am familiar with Franklin's book, in fact it is one of my two favourite books on the subject, besides the one by Peterson et al. (2011). In the latter, GARP is considered a presence/pseudoabsence method. Franklin is the one that "told" me GARP treats background and absence as the same.

My question is if about finding an algorithm which uses randomly generated points to sample broad environmental conditions (such as maxent and ENFA), and not to guess at absence locations, which GARP does when creating its rules.

In biomod you can choose whether the pseudo-absences are random or any other of 3 strategies which are available. But, of course, you can check the biomod2 manual.

I thought you would know the Franklin book, but I mentioned it anyway.

Cheers

Frederico

Any algorithm that is usually used for presence/absence data can be used in a presence/background or presence/pseudoabsence setting.  So this includes generalized linear models (GLMs), generalized additive models (GAMs), boosted regression trees (BRTs), random forests (RFs), classification and regression trees (CARTs), 2-class support vector machines (SVMs), and more.  Algorithms that don't use absences, pseudoabsences, or background sites are 1-class or profile methods like BIOCLIM, DOMAIN, or 1-class support vector machines.  Note that some people distinguish background points from pseudoabsences by defining the former as a set of randomly located point across the entire study region, and the latter being sites that have not been surveyed but are presumed to be absent of the species.  In this case pseudoabsences are drawn from places that are generally far from the species' known presences--the further away the less informative, but the more likely they will actually be absences.  Mathematically the algorithms treat absences, background sites, and pseudoabsences the same, but because they represent different things the actual models will differ between them.  Do note that Maxent should not be used with absences or pseudoabsences... of the lot it's different in that way.

I don't know what GARP actually uses--background points or pseudoabsences.  However, there's no "algorithm" per se that chooses background sites or pseudoabsences.  Rather, the program that runs the algorithm does it (i.e., the Maxent software or the GARP software).  So if you were to run any 2-class algorithm in R, say, where you can customize to your heart's content, you can decide if you want to use background sites or pseudoabsences by simply selecting points that represent either.  So I don't think you can say a particular algorithm uses only background sites (except for Maxent) or pseudoabsences.

Although maxent may be labelled as presence-only SDM, at default settings it will generate randomly a (relatively) large number of (pseudo-) absence points (5000 if I remember well); with presence points 1000; the default number of random absence points may be varied, helpful if you would enter >1000 presence points.

I know this is not what you asked, but please see paper by Yackulic et al. Methods in Ecology and Evolution 2013, 4, 236–243.

By the way, Maxent's default number of background sites is 10,000.

Just a note on maxent. It's not a presence-only model. It cannot works without background and/or pseudo-absences. the same goes for ENFA and basically all algorithms. the only real presence only SDMs are those based on Mahalanobis distance and Pearsons' Fit.

Totally agree with Adam B Smith. Maxent has been long considered a method specially adapted to presence/background data, but recent work has shown that the algorithm behind Maxent is equivalent to a Poisson GLM that approximates to a Binomial GLM (logistic regression) if the number of absences is high enough (see this post by David Warton and references herein http://methodsblog.wordpress.com/2013/02/20/some-big-news-about-maxent/). If we accept that Maxent correctly manage presence/background data, a logistic regression would do the same work if background data is large enough (e.g. 10000). Therefore, you can use Poisson GLM or logistic regression as presence/background methods in your comparison. You can find an example of this approach in my paper "Modelling species distributions with penalized logistic regressions: a comparison with maximum entropy models" https://www.researchgate.net/publication/236870185_Modelling_species_distributions_with_penalised_logistic_regressions_A_comparison_with_maximum_entropy_models

Article Modelling species distributions with penalised logistic regr...