1

I have insect traps data with several hundreds insects caught every day.

In this paper, authors say "Running the model using the number of mosquitoes per trap-night resulted in clustering of high probabilities of occurrence around individual collection sites."

How comes that they used the number of mosquitoes at a single individual trap location since so far what I understand from Maxent is that you only need three columns, occurence (species), x and y ?

How can I manage to tell MAxent to take into account the number of insects instead of their presence only ?

I am very new to MaxEnt and I really fail to resolve that.

2

Maxent is supposed to predict probability of presence due to the co-variates you are testing (usually environmental variables). It is a presence only model and is not an abundance model so I really fail to see how the paper you mentioned use number of mosquitoes (the link does not work).

One option is of course duplicating that presence point represented by an x and y coordinate as many times as you caught mosquitoes and manipulate the remove duplicates setting of MaxEnt. But I want to warn you from the get go that will mess with the validation calculation giving you a highly inflated accuracy. And of course it is kind of cheating to use duplicates in a correlative model.

  • 1
    I changed the link and it should work now; indeed I guess they duplicated the points since they display maps with or without duplicated data, but it seemed very odd to me to do so. Thanks for the answer ! – Cobactan Sep 23 '15 at 23:06
  • Thanks, I think a couple of things are concerning with their methodology. 1. the covariates they used seem highly collinear (just from the groupings) and 2. comparing a non-issue (using duplicates) with a no duplicate result does not add anything new. As in duplicates were never meant to be used in the analysis. Unless the coveariates they used are really high resolution (for example sub-meter) and in that case they won't be duplicates. – yanes Sep 23 '15 at 23:26
1

Have you tried posting this to the MaxEnt Google group? You might get a better answer there.

Anyway, MaxEnt is for presence only data - I don't recall ever reading a paper where count data was used. What they probably did was something like put several points very close together in order to simulate count data at a single location. For example, if you were using a vegetation category as a predictor in your model, and you had derived it from Landsat8 imagery, each cell would cover 30 meters by 30 meters. You could easily put multiple points into a 30mX30m area, and it would give you a result (I'm guessing, haven't tested this) that was something like using point data. But, as you point out, this created erroneous results - i.e. spatial clustering of high probabilities. I don't think MaxEnt is built to handle probability distributions that arise from point counts. You may look into Point Process Models - Renner and Warton recently wrote an excellent paper on the method, and have an R package for it called PPM http://www.inside-r.org/packages/cran/spatstat/docs/ppm. There are a number of good tutorials out there to help you with it too!

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.