Study Region

[bkara...@gmail.com]

Hello, 

I have a question regarding study region. Suppose the below square represents a country. The left and right areas are different regions. I only have occurrence points from the left region, however I highly suspect that the species also distributes in the right region. I would like to model the distribution of the species to predict where can I likely find the species in the whole country (right region included). I also would like to project to future. In this study, do I take my 10000 background points from the left region only, or the whole square? 

[Jamie M. Kass]

If you have incomplete sampling that you are quite certain is incomplete, I would say you should take a conservative background sample from only the areas surrounding each occurrence point you have. You can do this by buffering them by some distance (ideally informed by dispersal capability, etc.), then masking your rasters to the buffers before you take the sample. I suggest this in order to avoid including areas that may be suitable and inhabited but unsampled, as this would give the model a false signal. You should also try and remove areas from the background that are suitable but your species is unable to disperse to, for the same reasons.

Jamie Kass
PhD Candidate
City College of NY

[bkara...@gmail.com]

Thank you Dear Kass,

I have been fiddling with it for a while now and it seems I have a dilemma in my hands. When I choose the background points from a limited area (minimum convex polygon) around the occurrence points and use all 19 bioclim variables, the prediction of the resulting model into the region I am intersted in ais ctually ecologically sensible. However if I use a subset of the bioclim variables (taking into account the correlation) predictions of the resulting model doesn't really look right and I have tried various different combinations. What would be the best route to take from now on? 

8 Şubat 2018 Perşembe 12:04:25 UTC+3 tarihinde Jamie M. Kass yazdı:

[Jamie M. Kass]

Sorry I’m so late to reply to this. Maxent often deals with correlated variables pretty well as long as the regularization is high enough — this will trigger a stricter penalty of complexity and result in variables getting tossed from the model that don’t do a good job of explaining the signal. Therefore, if you have no expert opinion about which variables to keep, it’s likely best not to toss variables indiscriminately based solely on correlations, and just tune the Maxent settings with all the candidate variables instead. I know that some will disagree with me, but what I just explained is more of the machine learning mindset.