Replicates and final model

[fishe253]

Hello,
I just posted a question that is related to this on ResearchGate and the Google MaxEnt group so sorry for the barrage of questions. I am trying to generate some ENMs for bats of North America, 33 species, and I am using the ENMevaluate function to tune the models. I am using maxent.jar and block folds with the background partitioned as well. From what I can tell, when I then filter through and select the optimal fc and rm combination, looking at the prediction for that model and other things like the variable importances, those are from a whole model - no folds and all my occurrences and background points. Is that right? I can see that the metrics like AUC, CBI, etc. are the average over the folds, but the model is the whole model. Is the idea that the cross-validation in the tuning portion serves as the validation metric? Just trying to see what to do next after the tuning and whether to use the whole model or an average of the folds as my final model.
Anyways, this also led to an issue I noticed in my results. When I run ENMevaluate, in the end, in my output folder, it just gives me the results and details for the last fold. I've tried turning on replicates and entering a number there, but it then turns my spatial folds into randomkfolds. It may just be something in my code, but just wondering if there is a way with spatial folds for it to save the data and results from each replicate. Below is the code for that part. Thanks for taking a look!

tune.args <- list(fc=c("L","LQ","H", "LQH", "LQHP", "LQHPT"), rm=seq(1,4,0.5))

oargs <- list(path="C:/Users/makafish/Desktop/NA bats ENM final materials/Results/Final_enmeval_Results/", validation.bg="partition", other.args=c("fadebyclamping=TRUE","outputformat=cloglog", "addsamplestobackground=FALSE"))

e.mx <- ENMevaluate(occs = coords, envs = vs_for_model, bg = bg_coords,
algorithm = 'maxent.jar', partitions = 'block',
tune.args = tune.args, parallel = TRUE, other.settings = oargs,
doClamp = TRUE, raster.preds = FALSE)

[jamiemkass]

Thanks for the question, and great running into you at Entomology 2025. So the summary is that:

1. I recommend that for most applications, we select the model fit to the full dataset for further analyses, and we only use the validation models for evaluating the model performance. So model tuning will tell you how accurate the model is at predicting the withheld folds, and then we use that information to select a final model built with the full dataset. Otherwise, we would be combining validation models, all of which are built with subsets of the training data, and some of which might have very poor performance (especially for spatial cross-validation).
2. There is no way currently to save the models and raster predictions from the cross-validation runs. If you really wanted to, you could tweak the code in tune.enm.R to save them. But as I mention in 1), I recommend against combining them for your final model.