Interpretation of habitat suitability values

[Olivia]

Hi there,

I recently used Maxent to examine habitat suitability in relation to a nesting bird species across Ireland. We had 280 nest locations, and used habitat and elevation as the environmental variables. The analysis included Jackknife to measure the importance of the habitat variables, and the model was run multiple times to examine the variability in the model. For this analysis the model was run 15 times, consistent with guidelines presented in Phillips et al. (2017), and the average model presented.

So the AUC values from each of the replicated samples were high, and with low variability (0.9076 ± 0.0148). However, when the probability values at each nest and territory was extracted, a relatively high proportion were in areas identified as of very low suitability. We used the approach of Bosch et al. (2014) to explain areas of suitability and based on the modelled probability values:

1.       Very poor suitability (0 – 0.5) = 93 nest/ territory locations (41% of locations)

2.       Low suitability (0.5 – 0.6) = 32 nest/ territory locations (14%)

3.       Medium (0.6 – 0.7) = 58 nest/ territory locations (25%)

4.       High (0.7 – 1.0) = 46 nest/ territory locations (20%)

While the resulting map does reflect well the key areas that the species would be expected to nest in, which is great, 41% of nests and territories (included in the input file) were in areas of very low suitability. On balance it looks like the modelled suitability map is not very useful. I wondered has anyone else had the same problem? And/ or how others might have applied the suitability ranges?

Thanks

Olivia

[Veronica Frans]

Hi Olivia,

Have you considered looking at different thresholds? Right now, you have 0-0.5 as low suitability, but this is kind of an arbitrary threshold of presence/absence. Presence/absence is not that intuitive with these 0-1 scores, so most maxenters use different statistical thresholds that are derived from model training, e.g., 10th percentile presence, maximum sum of sensitivity and specificity (max SSS). Anything below the average threshold from your 15 runs would be absence (or low suitability), and then you can use rank values above that for other levels of suitability. Hopefully the distribution of site scores will be a little different. And then judging these values against AUC would be more reasonable. 

I'm not sure which version of Maxent you're using, but at least for version 3.3.3.k, these values would be in the maxentResults.csv and in the species html that Maxent produces. And they would be the 'logistic' thresholds (since your values are 0-1) and not the cumulative ones.

Check out Liu et al. 2015 for more info:

https://onlinelibrary.wiley.com/doi/full/10.1002/ece3.1878

Hope this helps!

Veronica Frans

--
You received this message because you are subscribed to the Google Groups "Maxent" group.
To unsubscribe from this group and stop receiving emails from it, send an email to maxent+un...@googlegroups.com.
To post to this group, send email to max...@googlegroups.com.
Visit this group at https://groups.google.com/group/maxent.
To view this discussion on the web visit https://groups.google.com/d/msgid/maxent/2e383321-b97c-4d38-9050-5d5027074a5b%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Olivia Crowe]

Many thanks Veronica. That sounds like a good plan. Will give that a go.

Very best

Olivia

To view this discussion on the web visit https://groups.google.com/d/msgid/maxent/CAE%3D1uP4gQDi9nhi3zD8mkc9jKRdD-iAna3oUSRzrpBgfUfYuWw%40mail.gmail.com.

[Camilo Arias]

Dear all  !! I would like to ask what consideration will you make about selecting between two kinds of thresholds using only presence data: 10th percentile presence, maximum sum of sensitivity and specificity (max SSS). Will you consider the use of the max SSS, even when we are working with presence only data?  will you consider the one with the lowest Omision Rate? 

thanks in advance !! 

To view this discussion on the web visit https://groups.google.com/d/msgid/maxent/CABge%3DdCujSEk3%3DmT1YOLbECGBQpxNg5wn0awH-aMb6ZjqcGngw%40mail.gmail.com.

For more options, visit https://groups.google.com/d/optout.

--

Camilo Arias González

MSc - Biología Animal

Cel. 304 591 4750 Manizales, Colombia

[Veronica Frans]

No problem, Olivia. I also was thinking of another Liu paper that would be more helpful to you. Maybe you should read this one first. It explains more of the thresholds. 

https://onlinelibrary.wiley.com/doi/full/10.1111/j.0906-7590.2005.03957.x

Cheers,

Veronica Frans

To view this discussion on the web visit https://groups.google.com/d/msgid/maxent/CABge%3DdCujSEk3%3DmT1YOLbECGBQpxNg5wn0awH-aMb6ZjqcGngw%40mail.gmail.com.

[Veronica Frans]

Hi Camilo,

Check the following article, which I mentioned before in this email string:

Liu et al. 2013. Selecting thresholds for the prediction of species occurrence with presence‐only data. 

https://onlinelibrary.wiley.com/doi/10.1111/jbi.12058

According to their work, there are three threshold methods that aren't affected by pseudo-absence data for several SDM algorithms: maxSSS, prevalence, and mean predicted values for random points. 

On an applications side of things, when it comes to SDMs that lead to actions/decision-making, I consider a strict threshold (10th percentile training presence; tends to have cutoffs at higher values in the suitability range) or a more conservative one (max SSS; tends to have cutoffs at lower values in the suitability range). It depends on how much you know about a species. A more conservative one would be a safer choice, because it would include more areas, as not to miss potentially important ones. But having two results of two thresholds can also help to decide between areas, as well. I also have used these two thresholds in particular because I've seen other conservation/management papers use them.

In the end, the decision is ultimately yours. Also keep in mind that you don't always need binary outputs. They are just guidelines. Another cool way to use them is for your color scheme. 0 to threshold is one color, and then threshold to 1 goes along a gradient. Then the relative probability of presence is more clear.

Hope this helps!

Cheers,

Veronica Frans

To view this discussion on the web visit https://groups.google.com/d/msgid/maxent/CAJ%3Da1QsoaE9nE%3D8jz6_mJJ0hSfRMeUDVqM-cR78nEX45QWA3sA%40mail.gmail.com.

[Camilo Arias]

Hi Veronica !! Thanks a lot for your kind considerations regarding threshold selection.

May I ask you a big favor? !! Could you  send me the other conservation/management papers that have used those thresholds in SDM please !!!   

In my approach, i´m using maxent to assess climatic suitability for 8 endangered species, in the present period, and for different scenarios of climate change.

I use this climatic suitability to asses it´s relative % of protection through Protected Areas Network, so the selection of threshold is quite relevant. 

I have considered those two : 10 Percentile Training Presence, and maxSSS, and i have seen that for each species, results are diferent on wich threshold gets a lower omision rate. In some cases Sp1, the  10%TP gets lower cutting value, and lower Omission rate, but in another case Sp2,  MaxSSS offers results with lower cutting value (0.294) and lower omission rate (4.13%) than the 10%TP ( 0.328 - O.Rate 8.3%)

As a threshold selection is a species by species decision, what do you think about selecting for each species,  the one with the lowest omission rate ? 

thanks in advance !!

Cheers, Camilo.  

 

To view this discussion on the web visit https://groups.google.com/d/msgid/maxent/CAE%3D1uP7uDqtTzBgWxbCgJSGHe%3Dhq_bJHhKFsRYWBa2%2BZ8DJZ3g%40mail.gmail.com.

For more options, visit https://groups.google.com/d/optout.