Hi.
When I change Max numbers of background points of MaxEnt from 10,000 to 20,000, my AUC progresses from 0.89 to 0.92; Please guide me, In which condition can I change 10,000 background points of MaxEnt to 20,000? I have about 1500 records and my study area is southern Iran.
Anyway, If I use 20,000 background points, my analyses will be incorrect?
Please can anyone help me?
Hi,
This paper could help you to answer your question, especially section " A. Background data" :
Article A practical guide to MaxEnt for modeling species' distributi...
Good read,
Matthieu
If your presence point records represent incidental observations, the number (1500) is rather large and may be reduced by filtering (see our recent ASF China paper) to minimize spatial autocorrelation. In case you reduce the number of presence points to about 150 points, the default pseudo-absences (10.000) may yield a similar AUC as the 20.000.
The larger the number of presence points, the larger the number of background pseudo-absence points seems to make sense from first principle. The default (10.000) may have been set in the software on the safe side for about 100 presence points.
Hi
Dear Hein my 1500 records are not incidental observations. They are true records.
Anyway, When I change Max numbers of background points of MaxEnt from 10,000 to 20,000, my AUC progresses from 0.89 to 0.92;
The only way you can say "the increase in the number of background points actually leads to an increase to the accuracy of the model, is to test the model on an independent presence-background evaluation data.
This is because AUC is very sensitive to the number of negative values (background here). The same model can give you a much higher AUC if you just increase the number of negative values in your test dataset (see Sofaer et al., 2019). So, I assume this would be an inflated measure of your accuracy.
Some researchers showed that the number of required background points is much higher than 10,000 (Maxent default) e.g. see Renner et al., 2015. But, your case can only be justified if you use an independent presence-absence test/evaluation dataset.
However, I still recommend using a higher number of backgrounds, but use another test than the AUC reported by Maxent to evaluate the predictive performance of your model.
I suggest to consider the article by Renner & Warton:
https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1541-0420.2012.01824.x
From my experience about 20% of the cells which are neither NA nor holding a presence is a good starting point.
Hi Taghi
Backgrounds (bg) points inform the model about the environmental conditions of your study area. How the number of bg points affect your predictions depends on the spatial scale of your study area, both in extension and resolution (grain). A high number of bg points will cause overfitted results and a low number of bg points will return too generalistic results with poor prediction accuracy.
For example, if your study area is a raster grid of 20000 cells an you use 20000 bg points in your model your AUC will be close to one (by definition AUC cannot be =1), because the model is predicting exactly your presences/absences records. This result has poor prediction power because of overfitting, although AUC evaluation will be perfect.
You can try to adjust other parameters in the calibration of Maxent such as regularization or features (see Merow et al. 2013) and also combine AUC with other evaluation methods.
Finding the 'best' model is a complicated task that depends on the species, the study area and the objectives of your investigation.
Hope this helps
Benjamín Strappa Hi,
I like your answer to which I have one a clarification to seek from you. So, in the example you have given, 20000 cells and 20000 bg points is overfitting which makes sense to me because bg points coverage is 100% which can influence the result. In this line, is there any specific paper or method or rule of thumb (70/30, 60/40 %) which can approximately tell us this proportion or ratio of presence to bg points is good enough (generally for large mammal if it is animal specific)
Thank you.
Hi Sujata Sunuwar
to my knowledge there's no rule of thumb, because each case is different in terms of available data. I suugest you to try and evaluate different parameters, not just bg points (see Merow et al. 2013). If you read spanish you can see an example in my thesis: Thesis Mapa de hábitat para la conservación del huemul (Hippocamelu...
Some examples of different model evaluation appear also in Radosavljevic and Anderson 2014 or WARREN and SEIFERT 2011. Hope it helps.
Hi Benjamín Strappa
Thank you for clarification and it is unfortunate that i cannot read spanish otherwise I am definite that I could have learnt from your paper.
Thank you.
- Badges
- Science topic
- Similar topics
- Counseling Psychology
- Confusion