Dear colleague

My advice is to use maxent (free software) for your intended predictive spatial modelling. Not only have we used maxent on herbaceous plants (as in the attached), but also for an endemic tree species in the same area with interesting results. We predicted the occurrence of the tree in the NP, where none of the local botanist had seen the tree.

In my experience the hard work in spatial modelling is not so much in the modelling software but in the pre-processing of the predictor layers.

regards

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

Thank you. I have read about maxent and thought it sounded like the most appropriate software.

Hi Harley,

To achieve the best results you should familiarise yourself with a few common issues, namely spatial autocorrelation, multicollinearity, and sample selection bias.

There are a wide range of modelling algorithms available for presence-only data, of those Maxent is the most popular as it's robust and fairly simple to work with.

Alongside that it is worth reading into:

• Environmental Niche Factor Analysis (ENFA), available in the Biomapper software package (see: Hirzel et al. 2002).
• Boosted Regression Trees, available in the R package dismo (see: Elith et al., 2008).
• Inhomogenous Poisson point process, available in the R package spatstat (see: Warton & Shepherd, 2010).

Good luck!

Elith, J., Leathwick, J.R., Hastie, T., 2008. A working guide to boosted regression trees. Journal of Animal Ecology 77 (4), 802-813.

Elith, J., Leathwick, J.R., 2009. Species distribution models: ecological explanation and prediction across space and time. Annual Review of Ecology, Evolution, and Systematics. 40, 677-697.

Hirzel, A.H., Hausser, J., Chessel, D., Perrin, N., 2002. Ecological-niche factor analysis: how to compute habitat-suitability maps without absence data? Ecology 83 (7), 2027-2036.

Warton, D.I. & Shepherd, L.C. 2010. Poisson point process models solve the ‘pseudo-absence problem’ for presence-only data in ecology. Annals of Applied Statistics, 4, 1383–1402.

Maxent is a good method, and also a tool, freely available, see http://www.cs.princeton.edu/~schapire/maxent/

Maxent is easy to use, even though it has some issues. For instance, it is a so-called presence-only methods, i.e. you do not need absence data. However, the program creates its own absences (called pseudo-absences, because you do not really know that they are true absences), and you need to be careful then, limiting the area that it is allowed to pick pseud-oabsences from. I think the default in 10 000 pseudoabsences, and it is drawn from the extent of your model maps. So consider this before you start.

Maxent is scary to use in “neck-down” analyses, which I believe is common. You put everything in there, you keep all the default settings and you just hit the “run” button. And it looks great. But you should know that Maxent creates all kinds of interactions between you input variables, not regarding the ecological relevance of the interaction and without presenting you with them, if you do not look in the log file.

I think the question of that you want to answer and what you want to use your model for is very important. Do you want to understand (and visualize) general patterns? Do you want to develop the best possible map of your area? Do you want to transfer you model to a different area or a different time? Do you want to improve sampling design? I would say that we have three different purposes: 1) Ecological response modelling, in order to model relationships in order to find the general patterns in the overall ecological response, 2) Spatial prediction modelling, in order to optimize the fit and 3) projective distribution modelling, in order to transfer you prediction to a different spatiotemporal setting.

BRTs are also quite popular. I would also suggest checking out SAM ("Spatial Analysis in Macroecology"), which is a program for spatial statistical analysis. http://www.ecoevol.ufg.br/sam/

I you are familiar with R, Maxent can be run easily from there, http://cran.r-project.org/

The scary part of distribution modelling is that to a manager, a really bad model can look just as nice as a good model, and a colorful map is always very popular to use and the probabilities and insecurities are often difficult to communicate.

Good luck

Thank you all for your input. Trine, you made a good point. One that I had in the back of my mind. That I may produce a model that looks good but is absolutely rubbish.

I have hundreds of records of this tree and a good feel for what variables affect its distribution. Is there such a thing as too many records?

I would add one more thing to the previously said. For me the most important thing in species distribution model is the  knowledge that you would have on your target species or group. At the end all the modelling methods will give you a result, so the challenge is to assess if the results have biological sense. For me is mandatory to have expert kwnoledge on the species that your are modelling, this expert knowledge will let you have a good output. Expert knowledge will let you select the appropiate predictors, those that make biological sense for your target species.

Many renamed scientist with strong quantitative background use to run SDM using the commonly used environmental predictors such those available in worlclim, but they do not care on other predictors. In my experience an important predictor for neotropical raptors is canopy height and terrain slope, this aspects are not considered in other studies modelling raptors distribution.

About the number of records, the minimum number of records depend on the software. MaxEnt provide an acceptable prediction with at least 10 records. It is expected that with more records, the more accurate is the prediction. Pay attention to spatial autocorrelation among your records, you can explore if your presence records are auto-correlated, if they are you only need to choose one of a pair-wise autocorrelated. 

Finally keep alway in mind a rule in any geomatic-based analysis, the GIGO rule: Garbage In, Garbage Out. With this I am saying that softwarse will give you numbers, results, but depending of the input data the result will be or not useful.

Well, my suggestion is to use R to perform your SDM's. Is easy, free and if you use also RStudio... is "friendly"!

The collegues above have explain the main things that you must consider. The SDMs are increasingly applied as predictive procedure for purposes of conservation planning and management of ecosystems (Elith and Leathwick 2009). I can only to add some suggestions:

The most traditional SDM's are perform using a biological variable (for example the occurrence of a species) as response variable and a long serie of predictors, to explain that occurrence. The response variable can be 1 species (in SDMs), so a binomial variable (1 or 0 values) and the predictors most used are related to proportion or percentage of lad-cover typologies in around the sampled site + other environmental characteristic. Rememeber to test if are needed transformations of any type. For example can use BoxCox procedures to determine the adequate transformation. I thinkh that is in the library "MASS".

The GLM procedures are fine. Several good papers can be useful to build models (Guisan & Zimmerman to start...)

Furthermore is important to test for possible problems realted to spatial autocorrelation of the data (if your sampled sites are too much aggregate). A Mantel test is a good option...Mantel measures the correlation between two matrices typically containing measures of distance. You can use the library "ade4" in R.

The last considerations are about the new frontiers in SDM's: the consideration of the importance of heterospecific interactions or biotic interactions. As highlighted by Kissling et al. (2011) the biotic interactions have widely been ignored in species distribution models. Ecologists starting now to pay more attention also to biotic interactions (not only predation or parasitism) that can constrain the goodness of fit of traditional SDM's build only on "pure" environmental variables.... ;)

Good science, good job!

Further to Alex and Harley

Some of the issues raised in using SDMs are dealt with in the attached (multi-collinearity; sample/observation bias); further we are transparent about our predictor and model selection, rare in the literature to the best of our knowledge.

Article Where the bears roam in Majella National Park, Italy

Interesting point, Harley. Can there be too many point observations? From my papers (attached above) you can see that as little as 10-15 points can be sufficient across a range of species and spatial resolutions. But I have not come across research showing either the-more-the-better, or alternatively, a normal (Gaussian) distribution of the number of points versus a SDM performance evaluator. Would seem a valid research question. The answer may depend on the extent of the research area, the spatial resolution and so on. As you seem to have a lot of observations, you may obtain a first answer by randomly selecting a subset (say 100 out of 1000) of your observations and compare the resulting SMDs. My null hypothesis would be: more is not necessarily  better.

Hein point is good. Harley, have a look on this http://onlinelibrary.wiley.com/doi/10.1111/j.1365-2699.2006.01594.x/abstract?deniedAccessCustomisedMessage=&userIsAuthenticated=false

Maxent is the top performer, but beware that you will have trouble publishing your results because of the above-mentioned issues. Read Jane Elith's recent articles on the topic for more information.

If you are familiar with R, I would recommend the "biomod2" package, which also has some nice tutorials.

For accessible texts into species distribution modelling, I'd recommend the following:

Pearson: http://perceval.bio.nau.edu/downloads/grail/climate_seminar/section2/pearson07.pdf

Franklin and Miller: http://books.google.com/books?id=PyhQV3o6WngC&printsec=frontcover&dq=franklin+miller+mapping+species&hl=en&sa=X&ei=-IgRVNXDLMSAygSh24D4AQ&ved=0CB0Q6AEwAA#v=onepage&q=franklin%20miller%20mapping%20species&f=false

In addition to the useful suggestions above, I would also recommend carefully thinking about what ecological processes/variablesthat  should be important for the species you are trying to model (lots of papers by Mike Austin on this subject). Some of the tools mentioned above, as well as many others (e.g., random forests), can be incredibly flexible and fit data well, but may also perform poorly when predicting responses elsewhere due to overfitting.

though I am late, I will send some answer. There are several techniques to limit the overfitting in Maxent modelling. Among these, I think that AIC/BIC evaluation is nowadays the best. Maxent in se doesn't calculate AIC (Akaike Information Criterion) or BIC (Bayesian Information Criterion), but ENMTOOLS does it. You can download this here: www.danwarren.net/enmtools. Take a look at that, and try to use it with your data.

In my experience (mainly on bird data) Maxent+EnmTools work fine with datasets spanning from 10 to 1000 presence points. Beyond this limit it is necessary, to my opinion, to perform other procedures to limit overfitting problems and inexplicable relationships with environmental layers. A good rule of thumb is to limit modelling to linear and quadratic features, but much more can be told on this topic. If you are interested in working with maxent, I will send you further suggestions...

Dear colleagues, Helen and David in particular.

Publishing SDMs generated with maxent is in my experience not more, or less, difficult than for GLMs (see my RG contributions).  I choose GLM for deforestation and forest expansion because reliable presence/absence data (satellite imagery) are evidently available for such research objects. For flora and fauna over larger areas more often than not only presence data are feasible and/or reliable.

Over-fitting and its resulting lack of SDM transferability may be an issue for both machine-learning (e.g. maxent) and GLM type of SDMs. Transferability depends in my experience, among others, on the selection by the modeller of the number and type of environmental predictors, also in relation to the spatial resolution of the model and the original environmental data (see my findings on the soil variable in Majella). In the literature, both the number and type of predictor variables seem mostly selected for convenience without any data quality consideration (e.g. meteorological grid data are mostly untested extrapolated point data). This weakness applies to the scene setting publications on SDM as well.

In my first response to Harry I referred to our endemic tree case. In fact we tested GLM versus maxent with the same tree occurrence data and environmental predictors and found no difference in model performance; variable, spatial resolution and model selection may be more decisive than the chosen algorithm. Although indeed in biomod you can test contrasting SDM algorithms. My prediction? Out of ten, GLM is the best performer for presence/absence and maxent for presence data only.

have fun

Thanks everyone for your contributions. As I first thought, this is a very large topic I am about to embark upon and I have a lot of reading to do. Even this thread offers a good amount of background. But please don't let that stop more people from contributing. 

Very informative discussion. Thank you all