Odd difference in the response curves between ENMeval, biomod2, and MaxEnt GUI runs

[yucheols]

Hi Jamie,

I wanted to reach out and hear your thoughts about a very weird difference in MaxEnt response curves I observed between ENMeval, biomod2, and MaxEnt GUI runs. There is no issue with the ENMeval package itself, but I just can't wrap my head around what might be causing this difference.

So, our team is modeling the habitat suitability of a marine organism across the Pacific. We used a global scale raster data of marine environmental variables downloaded from Bio-ORACLE.

First, we ran MaxEnt GUI in default settings as a test run. The cloglog output and response curves corresponded well with our initial predictions based on the species' ecology, with the y-axis values ranging from 0 to 1. So, we moved on to biomod2 to generate ensemble models using BRT, RF, and MaxEnt.

When we looked at the response curves of the biomod2 runs, however, we saw that the maximum y-axis values of the curves were extremely small (< 0.000003). This was indeed very confusing. So, we next ran MaxEnt by itself in ENMeval with various parameter combinations, as well as the auto-features used by the initial default GUI run (which was LQP 0.05). These runs also returned response curves with extremely small y-axis values.

Below are the response curves output from the ENMeval run, as well as a response curve of one of the input variables from MaxEnt GUI run:

Different curve shapes aside, I'm having difficulty understanding why the x-axis and y-axis ranges would differ like this between the MaxEnt GUI and ENMeval because the input data were identical between the two runs. I've never encountered this kind of issue before and feel very much stuck getting to the bottom of it.

Do you have any thoughts on what might be the root cause of this issue, or have you encountered similar issues before?

Thanks in advance for your help!

Best,
Yucheol

[jamiemkass]

First off, thanks for using the package. biomod2 uses ENMeval to tune and run Maxent models, so that might explain the low values in both pkgs, but if you made an ensemble and still got small values, that would mean RF and BRT had small values too?

Regarding those values, it's a mystery to me too. There can be differences between the output of maxent.jar runs and ENMeval tuning exercises because of some automatic occurrence (and background) cleaning in ENMeval, and possibly in the evaluation stats due to the way background records are treated in spatial cross-validation (see the documentation), but I haven't seen cloglog prediction values so low before.

My only idea is for you to send me a link to download a sample of your data so I can investigate.

Best,

Jamie

[yucheols]

Hi Jamie,

Sorry for the late response! I emailed you the dataset.
Thank you for your time!

Best,
Yucheol

[jamiemkass]

Okay, this required kind of a deep dive. So turns out, and I didn't know this, maxent.jar plots response curves differently from the R functions dismo::response() and predicts::partialResponse(). Both set x to the range of the variable in question and change the other variables to their means, then use the model to predict the data frame. However, maxent.jar uses the occurrence record means, and the R functions use the occs + background means. Because your occs (which are coastal and very low bathymetry) have environmental distributions that are very different from those of the background for multiple variables (bathymetry and dts especially), the R functions are always making predictions for values of one or more variables that are way outside those of the occs. So they always show very small predicted values, while the maxent.jar response curves show high ones because they use the occ means. This is a kind of a special case for your data set because the environmental distributions for the occs are so different from those for the background for several variables -- usually the background means are no so far from the occs, and we probably don't really notice the difference between the maxent.jar and R function response curves.

Not sure which is preferable, to use the occs means or the occs + bg means, but neither is more correct than the other -- just matters what you want to show.

If you want to plot these kinds of response curves in R, you can simply build your own tables and plot them like this (kind of rough code). You can customize the plot how you like.

# make a list for your data frames
resp.ls <- list()
# get the means of all the variables
varmeans <- apply(occs.z, 2, mean) |> t() |> as.matrix()
# loop over the variables in turn
for(i in names(p.envs)) {
  # define a buffer for both sides of the response curve
  # I made this based on the sd * 5, but could be anything
  buf <- sd(occs.z[,i]) * 5
  # make the x-axis as a range of the variable in question
  x <- seq(min(occs.z[,i]) - buf, max(occs.z[,i]) + buf, length.out = 50)
  # make the data frame for your new prediction
  df <- data.frame(x, varmeans)
  # remove the variable in question from the varmeans table
  ind <- which(colnames(varmeans) == i)
  df <- df[, -(ind+1)]
  # set the column name to the variable name
  names(df)[1] <- i
  # make the prediction
  df.p <- predict(me1, df)
  # make a table for plotting the resp curve
  df.plot <- data.frame(x = df[,1], pred = df.p)
  names(df.plot) <- i
  # load the table into your list
  resp.ls[[i]] <- df.plot
}

# set the plotting env to fit your plots
par(mfrow = c(3,3))
# loop over the variables and plot each resp curve in turn
for(i in 1:length(resp.ls)) {
  plot(resp.ls[[i]], type = "l")
}