Some researchers use single-algorithm (for example Generalized Linear Models, Maximum Entropy, Generalized Additive Models, Neural network, Random Forest, etc.) while others are more willing to work with ensemble projections and consensus models (i.e. mean or weighted mean of more than one algorithm). It doesn't really matter; you can choose what you like best, so long as you describe the advantages and limitations of your model and do a sensitivity analysis.

This one uses Maxent:

Using species distribution modeling to delineate the botanical richness patterns and phytogeographical regions of China, Scientific Reports, Nature, 6, Article number: 22400 (2016)

doi:10.1038/srep22400

Dear Jorge,

For presence only data, my personal preference is Maxent. There are plenty of references which point to the performance of the algorithm (more than a few comparative studies of performance). An ensemble approach would be good too, but overall I think what is probably more important than choice of algorithm is the underlying bias within the data itself. GBIF data, whilst useful, is riddled with errors and how you treat the bias will most certainly effect the usefulness of your model regardless of algorithm (there is a nice paper by Fourcade et al. 2014 on bias). As Andrew pointed out, a sensitivity analysis would be welcome (try Caret package in R).

The spatial scale will also greatly effect how useful your final model/s are. Scale will be relevant to your question and your species/genera.

best regards,

Further to earlier answers as a user of GLM and Maxent, the SDM algorithm may be less important than the spatial and thematic quality of the presence data and the environmental data. On the latter, the completeness of the set of variables may effect your result more than the algorithm. For some examples see attached.

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

Article Where the bears roam in Majella National Park, Italy

Article Pinus mugo Krummholz Dynamics During Concomitant Change in P...

Thanks a lot, it seems that there is no difference on using algorithms and the most used and aproved is Maxent.

Jorge, there are many differences between Maxent and Bioclim, Bioclim and GLM... but there is not a best algorithm in all cases. On the other hand, the main tools to measuring the discriminatory capacity are frequently misleading. Think about the occurences of species (presence records), the environmental context, the ecology of species, the extent (study area), the predictors, etc. The performance and reliability of the algorithms depend of these factors.

I agree with the suggestions above that there is not one best method; different algorithms perform different under different conditions.

But it is important to keep in mind the purpose of the modeling exercise. If the goal is to model presence/absence or abundance in new (unsampled) locations or under future conditions it is important to be wary of the risk of overfitting. Model tuning by means of spatial or temporal cross-validation spatial can help. Instead of holding out random data points, with this method you leave out spatial (or temporal) chunks.

For an example see this paper

Article Consensus forecasting of intertidal seagrass habitat in the Wadden Sea

Furthermore, if you use predictors highly correlated and the default parameters of Maxent, your model will be increased risk of overfitting and overparameterization. But, paradoxically, your model will have a "good" (LOL) discriminatory capacity.

it's fairly easy to achieve a misleading measure of the performance of discriminatory capacity: it only takes -among other things- to increase unreasonably the study area (extent). 

Sometimes, the goal of the research is to simulate an overprediction because we need to be very cautious about the spread of invasise species or the conservation of a rare species.

In the papers referenced in my earlier answer, you may have a look at various ways to exclude highly-correlated environmental variables. High correlation between climate variables, between climatic variables and elevation, between land use/cover and climate variables are the rule rather than the exception, obviously as we deal with an (eco-)system. However, pairwise comparison and subsequent removal are often arbitrary. More transparent and reproducible is the step-wise removal of the least contributing variable (applies to both GLM and maxent; see my references above). Even after exclusion of highly-correlated variables a high number of remaining variables may have an effect (e.g. 'over-fitting') on the resulting SDM. Maxent seems robuster than the other tested SDM algorithms (see our Phytophthora Andalusia/Australia SDM) for the number of variables.

Dear George,

I realize it 8 month has passed since you posted your question. But I think, you still may find my comments useful in your work.

The common feature of all models is that all are abstractions, which reflect only some portion of reality. I personally find it useful to use different methods complementary, to better understand, how the species utilizes available ranges of each eco-gegraphical variable. For instance a number of methods (FANTER, MADIFA, ENFA (Hirzel et al, 2002; Calenge, 2008)) allow to evaluate marginality and specialization effects of variables, which may add understanding of how some values are restricting, and to make a sensible choice in variable selection. I suggest you review papers on misuse of modeling to avoid some very common pitfalls (See Elith et al., 2011; Yakulick et al, 2012; Merow, 2013). In my opinion many distribution range modeling publications suffer from practitioners seeking for a shortcut ways to choose their variables. Therefore I personally prefer to carefully choose a limited list of eco-geographical variables that have interpretable ecological meaning, based on the known facts on the ecology of the target species. When it comes to geographic variables in general and climate variable in particular, the correlations among them is rather a norm. I think that may be acceptable to use correlated variables, as long as they do not overlap in terms of their ecological meaning. One can use histograms produced by AdehabitatHS, or Maxent response curves to understand how does the species of interest utilize the variance of the variable, to make further choice on the variable selection. Also, as it has been mentioned by others, the scale and geographic extent determine the variance of available values. One problem I often see is some authors combining several variables that essentially reflect the same pattern , and therefore introduce redundancy (like mean annual temperature, winter minimums and summer maximum). I find it more meaningful to choose either one or another. Different redundancy comes from common combination of Bioclim varlabes with digital elevation model (DEM). Bioclim variable rasters were originally estimated with use of DEM. As a result DEM and the majority of Bioclim variables are expected to correlate highly, especially at smaller scales.

Below are few papers you may want to review to start with:

Elith, J., Phillips, S., Hastie, T., Dudík, M., Chee, Y.E., Yates, C. 2011. A statistical explanation of MaxEnt for ecologists. Divers. Distrib., 17: 43–57.

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

Yackulic, C.B., Chandler, R., Zipkin, E.F., Royle, J.A., Nichols, J.D., Campbell Grant E.H., Veran, S. 2012. Presence-only modelling using MAXENT: when can we trust the inferences? Methods in Ecology and Evolution, 4, 3: 236–243

Hope this helps!

Tigran