You will see diverse answers. The "dogma" is that Maxent is the best, and there is no denying that it is a good tool. For this viewpoint, see Elith et al. 2006 (linked below). But for a more nuanced viewpoint, see the Qiao et al. paper, also linked below. All the best, ATP

Article Novel methods improve prediction of species’ distributions f...

Article No Silver Bullets in Correlative Ecological Niche Modeling: ...

You will see diverse answers. The "dogma" is that Maxent is the best, and there is no denying that it is a good tool. For this viewpoint, see Elith et al. 2006 (linked below). But for a more nuanced viewpoint, see the Qiao et al. paper, also linked below. All the best, ATP

Article Novel methods improve prediction of species’ distributions f...

Article No Silver Bullets in Correlative Ecological Niche Modeling: ...

Dear Sam et al,: Thanks, a quick comment perhaps on DOGMAs...(probably a lame discussion by now; sorry if so): Maxent is 'just' point-and-click, less of a science or  reflective, unfortunately fully bended and bound by a certain probability mind set, and by now driven by a 'clique' ,and it presents a cottage industry of sorts. So that dogma that Town mentioned is not so widely shared by more advanced modelers and for data mining and in complex data situations (exactly the ones as you have for the EU and beyond). So beyond maxent, I propose to read up more on ensembles, machine learning and coding for progress; here an example:

Ferandez-Delgado M, Cernadas E, Barrow S, Amorim D (2014) Do we Need Hundreds of Classifiers to Solve Real World Classification Problems ? Journal of Machine Learning research 15: 3133-3181.

Hint: the good and the best stuff is not always free...and/or in R; rather vice versa. Science and conservation always demands for 'the best' though, not to be conservative but precise, and be fully transparent and repeatable.

That's the goal.

I am so glad we are moving away from reductionism, or parsimony either way; and thanks for that!

Kindly

   Falk

Hi,

As Falk pointed out, MaxEnt is point and click, hence most popular, other thing is Elith et. al. figured out it was one of the good performers. 

Then people started using it pretty easily, I already quit it as I did not satisfy with the result from it when I tried in complex landscape of Himalaya. 

try to find other options, good luck. 

Hi,

Maxent is a presence-background method (only presence methods does not exist), and probably the most widely used Species Distribution Models algorithm. I would say it is an easy to use method, and as already said by Townsend Peterson I would spend some time to choose the most appropriate method/s according to your interests.

Although it is not updated, you can find some help in Franklin, J. (2009). Mapping species distributions. Spatial inference and prediction. Cambridge University Press. Cambridge. pp. 105-112, with a brief description of different methods.

Are your species widely distributed or do they have restricted distributions? In any case I would recommend you to take into account the prevalence of the species.

Best regards

I agree with AT Peterson, but I think there is not any "dogma" about Maxent, I would call it "trend". On the other hand, concerning what Antonio said, there are presence-only (P/A) methods like the environmental envelope methods or distance-based. In relation to presence-only methods, you can find Bioclim, Domain, Mahalanobis. Some modelling techniques like Maxent use background data for model construction but these are still classified as presence-only methods, but these should be called presence-background methods. If you have not confident about your absences data, I recomend you a presence-background (P/B) like Maxent. Maxent is pretty good to start...Then, you will discover the "buts". Nowdays, I use: Maxent (P/B), Random Forest (P/B), GLM (P/A), and GAM(P/A). 

Many thanks for the responses. I will take them on board. I might just start to play around with MaxEnt first to get a feel for it, since it seems to be the most popular and easy to use, then try a few others. I may be back with more questions as this project progresses.  Cheers

If you start using Maxent, at least, work omit the default parameters and experiment with different settings (features, outputs, random seed, random points, test percentage, regularization multiplier, etc). Working with a well known species helps to understand the algorithm and its predictions. Watch out for the trap: The Maxent GUI is dangerously easy to use.

References:

Elith, J., Phillips, S. J., Hastie, T., Dudík, M., Chee, Y. E., & Yates, C. J. (2011). A statistical explanation of MaxEnt for ecologists. Diversity and Distributions, 17, 43–57. http://doi.org/10.1111/j.1472-4642.2010.00725.x

Guillera-Arroita, G., Lahoz-Monfort, J. J., & Elith, J. (2014). Maxent is not a presence-absence method: A comment on Thibaud et al. Methods in Ecology and Evolution, 5(11), 1192–1197. http://doi.org/10.1111/2041-210X.12252

Halvorsen, R., Mazzoni, S., Dirksen, J. W., Næsset, E., Gobakken, T., & Ohlson, M. (2016). How important are choice of model selection method and spatial autocorrelation of presence data for distribution modelling by MaxEnt? Ecological Modelling, 328, 108–118. http://doi.org/10.1016/j.ecolmodel.2016.02.021

Merow, C., Smith, M. J., & Silander, J. a. (2013). A practical guide to MaxEnt for modeling species’ distributions: what it does, and why inputs and settings matter. Ecography, 36(10), 1058–1069.

Phillips, S. J., & Dudı, M. (2008). Modeling of species distributions with Maxent : new extensions and a comprehensive evaluation. Ecography, 31, 161–175.

Radosavljevic, A., & Anderson, R. P. (2014). Making better Maxent models of species distributions: Complexity, overfitting and evaluation. Journal of Biogeography, 41(4), 629–643. http://doi.org/10.1111/jbi.12227

Renner, I. W., & Warton, D. I. (2013). Equivalence of MAXENT and Poisson Point Process Models for Species Distribution Modeling in Ecology. Biometrics, 69(1), 274–281. http://doi.org/10.1111/j.1541-0420.2012.01824.x

Thanks Oliver for the advice and the links. I'll play around with the settings.

For a comparison of ten SDM algorithms see attached. In short, GLM and MaxEnt perform best in the defined context.

Important seems the range and quality of the independent environmental variables (aka: predictors/ co-variates). The good old garbage in, garbage out.. Particularly problematic are the widely, but uncritically used interpolated meteorological point data. Further, various observation biases for the dependent variables may play a role. On the latter three issues see the second attachment.

Article Where the bears roam in Majella National Park, Italy

Article Transferability of species distribution models: The case of ...

Many thanks for the article recommendations Hein.

...re. GiGo (Garbage in, Garbage out):

this is a very very relevant concept to test and to check though. We find in many (=most) cases that it is not true, hardly relevant. Many reasons why this is. In most SDM studies you find that even the most basic concepts of a classic research design and sampling (as per textbook) are violated...but ...it still works (as shown with other data, lines of alternative evidences etc). That's where the data mining and machine learning arguments I mentioned come in and can progress and outperform! Leo Breiman (inference through predictions). It does not always work, but if you can provide new information, new data and new insights, case is made. That's where the gain sits usually, and while maxent is pretty robust and good, it can often be outcompeted and boosted with other approaches in machine learning. It just needs a (science) mindset to explore and to test it.

So no, GiGo and such mindsets usually do not provide much new (it's circular and self-pleasing...) and should be challenged for sure. Time the textbooks are re-written... again. Keep me posted please by all means.

Thanks Falk.

With reference to Falk's hint that "the good and the best stuff is not always free...and/or in R; rather vice versa." 

Falk what good stuff would you recommend that is not always free?

Hi there, well, that's a good question, Roman: see it that way perhaps:

The modern sword (to win the battle...) is located in the best tool to answer the major questions of this world. Many of those are now linked with how to extract 'relevant' signals from (poor) data; all done in 'the best' way. Of course, machine learning is among those tools that can deliver that. Consequently a huge demand exists to use it. It's the underlying pressure and the driver to develop and to own it. Human nature pushes for it. As long as we have Capitalism allowed as the global rule (more or less) then there is a price on such tools. Certain dynamics unfold.  "The BEST" tools are not really in the Open Source domain. When it comes to data mining, see the Data Mining Competitions and who wins there and why. It's often commercial stuff which is locked away to most people. Whoever controls such stuff controls...the entire domain... and makes much money with it (Capitalism in the pure form).

I am happy to speak more about why R is not so much cutting edge when it comes to Machine Leaning analysis (hint: RandomForest implementations are not well optimized in R, either they are incomplete or mostly just repeating frequentist mindsets still)). More can be said. Let's agree that people much smarter than us have already worked on Machine Learning for decades, and such tools and approaches are beyond just 'point and click' (Maxent included). Lastly, science and conservation should not sell themselves short by using 2nd or 3rd class analysis  (parsimony being one of those). Nature conservation just deserves the best tools. We are not serious any other. My view. Kind regards  Falk

Ah yes, Faulk, so obvious now that you point it out that way; however, conservation is a poor competitor in a capitalistic environment -- maybe it has more luck in a democratic one -- but democracy is all too easily contaminated by capitalism, as we all know so well.

...sure, thanks! Considering that we only have one world, it's all finite, urgent time to think of a better economic paradigm then. Just to be serious

I propose, machine learning and conservation papers should point that out very clearly (we usually do that strongly, but in contrast you will very rarely find that in Maxent & Ecology publications, nor even the open access data they use and readily made available for others to see and to test. Often the funders and NGOs etc like to avoid the topic in full; fear & loosing them money).

Re. which software to use best: our own methods and work try to use those ones as much and as clear as we can (consider us widely unfunded). But by now, it becomes clear that the real culprit of environmental failure sits much more in the management and policy arena (regardless of the machine learning algorithms used); the global framework we mentioned (and just think of UNEP and such type of 'leadership'). Modeling is probably the least of our global concerns; for now.

Enough said, thanks and please follow up with email/phone outside as wanted. Kind regards  Falk

Morning !

a) http://www.ess.inpe.br/courses/lib/exe/fetch.php?media=cst-304-embio:refs:franklin_miller.pdf

b) https://www.researchgate.net/profile/Andrew_Peterson10/publication/230709994_Ecological_Niches_and_Geographic_Distributions/links/5860be4408ae8fce4903bc00.pdf

Two fundamendal books on the topic. What the previous scientists sais are correct - always keep in mind the GiGo (Garbage in, Garbage out), it applies to any aspect of modeling.

Best

D.

Book Ecological Niches and Geographic Distributions

On the practitioners level, a comparison between SDM algorithms including presence versus presence/absence is attached. Among the best performing SDMs differences are minor. Similarly, spatial resolution is of minor importance (attached). The range and data quality of environmental predictor variables and the delineation of the Area of Interest may have a larger impact on the results (attached).

Article Fine resolution distribution modelling of endemics in Majell...

Article Where the bears roam in Majella National Park, Italy

Article Transferability of species distribution models: The case of ...

..whatever you do, and whatever the question: Machine Learning has over 100 algorithms, and I propose to pursue that the most, beyond just 'falling in love with one model' algorithm (whatever that algorithm is). By sticking to GLMs, or Maxent etc you just show your ignorance of the topic (ensemble models as where the power really sits), and lack progress for conservation management (in case that matters). Rule of thumb on that subject "many weak learners make for a powerful single learner"! That's my suggestion here. 

Dear Samuel,

I prefer the packages 'dismo' and 'biomod2' of R statistical software, but they needs some R programming skills. With dismo, you can build Boosted Regression Trees (BRT, Gradient Boosting Model, GBM) which deals with presence/absence data and is one of the best performing species distribution modeling (SDM) methods. It is a kind of machine learning algorithm based on a lot of small trees (CARTs). In contrast to Random Forests (RF), it uses boosting instead of bagging.

With biomod2, you can build model ensemble. The package implements the following SDM methods: GLM, GBM, GAM, CTA, ANN, SRE, FDA, MARS, RF, and two versions of MAXENT.

A user-friendly, but less sophisticated alternative, beyond Maxent, is the Bioclim algorithm implemented in DIVA-GIS. It performs climate envelope modeling sensu stricto, which is an easy-to-understand and easy-to-interpret SDM method on presence/pseudo-absence data in contrast to Maxent, which is sometimes treated as black-box method.

Please find attached some links related to dismo, biomod2, BRT and DIVA-GIS. I attached also a comparison of SDMs in terms of their ability to deal with multicollinearity of the training predictors, which you may find interesting for selecting an appropriate SDM method for your research.

HTH,

Ákos

http://www.diva-gis.org/docs/DIVA-ex-2-peanuts.pdf

https://cran.r-project.org/web/packages/dismo/vignettes/sdm.pdf

http://www.diva-gis.org/download

https://cran.r-project.org/web/packages/dismo/index.html

https://cran.r-project.org/web/packages/biomod2/index.html

Article A Working Guide to Boosted Regression Trees

Article Collinearity: A review of methods to deal with it and a simu...

Many thanks to everyone who has provided comments and advise. This is obviously a very topical area. All of the comments have been, and will continue to be useful, not only for me but for other followers, so please keep them coming. It is definitiely useful to re-read over them.

What's peoples experience with the R package of MAXENT?

HI

Just to add to Ákos Bede-Fazekas point, it is important to note that BIOCLIM will give you a binary presence/absence output, with no gradient of suitability for the species across the landscape. I don't have a link but particularly if you are modelling under climate change it is best practice to use the aggregate of multiple algorithms, as can be done in biomod2 packaqge in R, or the relatively new SSDM package in R. See Araujo and New (2007) and Thuiller et al. (2009) for the benefits of using an ensemble approach.

What is your aim? To explain or to predict? Maxent is not as good as it was claimed, and not that bad as can be heard nowadays. Tons of ink, and pixels, dealing with this topic.

Have a look to this paper:

http://abc.museucienciesjournals.cat/volumen-39-1-2016-abc/testing-the-efficacy-of-downscaling-in-species-distribution-modelling-a-comparison-between-maxent-and-favourability-function-models/?lang=es

Bestregards

well, looks initially like a nice publication and Special Issue, but why not dealing with the REAL problems instead, like the EU, capitalism, neoliberalism, open access data shared and how to make it all better ? This work Antonia-Roman Munoz cited is a typical example of inefficiency for conservation, and just counting/modeling the deckchairs on the Titanic . It's hardly science/reflective! You can get lost in modeling endless, ignoring the wider framework of governance and context, but it's usually a bias and ineffective, certainly when single algorithms get involved. In Machine Learning, you have 100+ algorithms handy; so why not using them and ask relevant -more holistic- questions ? An AUC question of "82 or 87%" hardly matters when species are lost left, right and center. Now what ? =>Get with the good program guys, unless you want to be overtaken by 16 years olds https://www.youtube.com/watch?v=VFkQSGyeCWg !

It depends on your research question. Are you looking for distribution only or trying to predict the distribution? To predict the distribution MaxEnt is best but you need to be extra careful in selecting the variables for modelling.

Read Phillips et al 2004 and references cited in the paper.

Dear Shaheer Khan, thanks, but in all due respect, I am pretty sure from your reply that you are not on top of the latest and relevant literature and facts and software. Perhaps one may agree that Maxent was a tool forward and robust (better than any AIC and p value linear regressions); sure, but if you put effort into it, it's easy to outcompete and to further optimize, in most cases. Here two reads for you, those are pretty old by now and more can and should be done; key issue remains 'science what for' and 'by whom' ? I hope you work it all with open access data sharing and open source code.

http://jmlr.org/papers/volume15/delgado14a/delgado14a.pdf

https://www.springer.com/gp/book/9783319969763