Dear Ivannalis,

To answer the first question, yes, it is acceptable to perform tuning in another package (in your case, SDMtune) and then input the optimal model settings into another package like dismo; many researchers perform tunning in ENMeval, for example, and input it on packages like sdm or biomod2. It is perfectly normal and acceptable to do that. For instance, I use ENMeval to perform model tunning first, explore the results of tunning until I'm satisfied with the best setting(s), then I take these settings to my code which is implemented in the sdm package.

About the background points, generating it randomly (or by another mean, like a climatic envelope, environmental distance, etc) to use in Maxent modeling is the norm since Maxent uses background data and not true absence and/or pseudo absence; this is something that the literature complicates a lot since some authors like to treat background data and pseudo absence as the same thing (i.e. as synonyms) when, in fact, they are different entities from an ecological viewpoint. Now, the behavior you described here is curious (or "weird", if you prefer) since the number should not be reduced, but as you said that may be happening due to block allocation, I would suggest that you investigate it further to try and make the number stay constant. If that's not possible, I would advise investigating tunning in another package, like kuenm or even ENMeval (you could try a simple tunning, using one feature class and one regularization, for example); biomod2 also has an in-built function to tune models before actually modeling the data. That said, regarding your second question, is it acceptable to reduce the number of background points? There is no short answer to this, not even a single answer I'd say (others may disagree with me). Background data (and pseudo absence, too) will affect model performance, will bias it (negatively or positively), and affect the final projection. There's so many things to consider during model tunning and training that there's really no simple answer. That said, here's what I would advise you to do:

1) have a clear objetive(s), understand what you're trying to do (e.g. "Is it a SDM or an ENM?", "Are you projecting to other areas or time periods?");

2) understand the limitations of your data (e.g. how many occurrences you have, how old are they, how are they distributed along your study area);

3) if you're using just one algorithm (i.e. Maxent), what's the best parameters you could tune for it considering your data? Are there just one set of parameters or you can tune two-three sets of different parameters that offer similar performances?

4) how you partition your data matters and you must decide which type of partitioning you're doing. That depends, again, on your occurrence data, on the objective(s) of the study, what's the best partitioning available considering how the occurrences are distributed in the study area, etc.

As a final advice, I would say that you should definitely look for dismo alternatives, like the sdm package (Naimi & Araujo 2016), and SDMtune alternatives if you want to explore other options, like kuenm (Cobos et al. 2019).

Best regards,

Gustavo R. Brito

Hi Gustavo, I really appreciate your comprehensive response, it made me realize that I still have a lot to learn about the fundamentals of MaxEnt. The insights you shared were very helpful. I now feel more confident using SDMs for model tuning. As for the background issue I raised, I’ll look into it further.

Best regards.