I am attempting to determine what the best species distribution modeling algorithm to use would be for a data set of species localities that comes from a regular monitoring program. More specifically, I have a data set that consists of about 40 sampling localities across a watershed that were all sampled in the same manner. For each site I therefore know the species richness (and what specific species are at each site). I want to use these data to predict what species are expected to occupy each of these sites given a limited set of environmental data and use this to analyze the mismatch between what species are observed and expected when using incomplete environmental data.

Are there any papers that utilize a similar approach?

  • 1
    Are you looking for a solution with a particular software? – ragnvald Jan 6 '17 at 19:41
  • I'm most familiar with ArcGIS and R so either of those would be best. The main thing I'm trying to identify is the method and I can figure out how exactly to do it later but if you have suggestions on both that would be great. – Kevin Jan 6 '17 at 19:57
  • Did you check this? – aldo_tapia Jan 6 '17 at 19:59
  • Briefly, but I will look into it further. From what I saw though it didn't really cover much about my primary problem which is that for many of the species sampled within the survey there are very low numbers of occurrences which would make it difficult to both train and validate the model. – Kevin Jan 6 '17 at 20:08

There is an old mantra in statistics: "your model is only as good as your data". In seeing your comment following the original question, it sounds like you are more interested in the power and effancy of your data than the actual model. It would be nice if you indicated what algorithms/models you are exploring and what outcome you desire. With this information, relevant recommendations could be provided regarding pros and cons of a given modeling approach. Some methodologies produce better estimates of 1st order variation whereas others produce 2nd order estimation that represents finer scale patterns. Statistically, both of these estimates would be valid.

I am not going to bother with singling out papers because, with a brief literature search you will find many papers comparing SDM methods, including a paper asking the question "do we need another SDM methods comparison". I agree with the paper questioning the need for additional comparisons. This is because I believe that we are asking the incorrect questions around performance and completely ignoring issues associated with expectations, data and extrapolation. Given your issue, this is exactly what you should be asking.

Following this opinion, I would recommend selecting a methodology that would produce sensible results and is easy to automate. This is because, ideally you could implement a Monte Carlo approach to produce many "quasi-random" estimates to quantify the spatial uncertainty in the estimate, say using the variance across replicates. Personally, in informal testing of MaxEnt (which I am not fond of and find somewhat invalid), given many runs of the model, the probability surfaces produced regions of uncertainty equating to +/- 1 standard deviation, which is worse than random and illustrates issues with spatial and statistical extrapolation of the model and potential instability in how the null is generated.

This brings up another issue in how the null is produced. For example, the MaxEnt software produces a fixed set of random or systematic samples. This is very unsatisfactory because there are no spatial constraints applied to the randomization. The model could perform very differently given a different null. This is something that could be explicitly explored in a Monte Carlo, as one of the simulation conditions.

The above point regarding MaxEnt stresses the need to narrow candidate models down to ones that meet your specific objectives, which can be variable in SDM's. Given a preference towards inference (or MaxEnt), I would direct one to Poisson Point Process models which provide a Bayesian framework to solve models using an MCMC. This would give one considerable inference around the results. For nonparametric methods the Random Forests and Kernel SVM models provide considerable power for binominal probability estimations. However, if one wants to explicitly incorporate spatial process into the model alternative approaches must be implemented adapting existing approach (eg., kernel weights in random forests) or selecting a model specifically designed for this purpose (eg., spatial autoregressive, conditional autoregressive, mixed effects models).

I would add that, since this data is produced via a return survey, it would be prudent for you to explore occupancy modeling which addresses observational bias and, in turn, adjusts probability of occurrence.

  • To clarify, the outcome I desire is to have a reasonable estimate of the species assemblage predicted to occur at each site given a set of habitat variables known to influence the distribution of said species which I can compare to the observed assemblage at each site. I then want to use this information to test hypotheses about several factors that may have an influence on the ability of certain species to inhabit a site (e.g. presence of a particular competitor) which would be evidenced by disagreement between the predicted and observed set of species. – Kevin Jan 6 '17 at 21:30
  • If your survey data is presence/absence for each species then I believe that an occupancy modeling approach would suit your needs. You can correct for detection probability and make estimates of occupancy probability in a multivariate model. It is, in essence, a fancy binomial logit model. I would recommend reading MacKenzie et al., (2005) "Occupancy Estimation and Modeling". There is the MARK software and a few R packages, including "unmarked", that support this type of modeling. If put in a Hierarchical Bayesian framework you could model all of your species simultaneously. – Jeffrey Evans Jan 6 '17 at 22:32
  • I would not feel comfortable designating this survey data as presence/absence. – Kevin Jan 9 '17 at 14:11

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.