Since habitat suitability is defined as " the capacity of a given habitat to support a selected species", by definition, all other things equal, your model will be more accurate if you do not show areas that you know are unsuitable in your modeled expression of suitable habitat.

Hi Shabnam,

It depends. If your study question is specific to only the Park’s mountain areas and/or you are extremely confident in the total absence of goats from non-mountain habitats, it might be perfectly fine to focus your analysis on just the mountain areas. However, if your study question pertains to all of the Park and goats could or do occur in non-mountain habitats (see comment below), data from all areas and elevations should arguably be included. When we (Hodge et al. 2017) were trying to predict habitat suitability for native trout, it was as important to include data from a reasonable number of vacant habitats as it was to include data from occupied habitats.

If elevation is indeed the best predictor of suitability for goats in the Park, model predictions should reflect that relationship. Incidentally, that lowland (i.e., non-mountain) areas were predicted as being suitable for wild goats suggests that goats do occur outside the mountains (or that the model isn’t working appropriately). If proximity to water sources is a less “weighty” predictor of goat occurrence than elevation, that too should be evident.

Anyway, I would include both elevation and proximity to water sources in the model. If you’ve already done that, or do so, and get the same unexpected results, there could be a number of possible explanations. I would be happy to offer additional thoughts/suggestions, depending on the scope of the study question and the results of any subsequent model runs.

Good luck, and hope that helps!

Brian

If you are completely sure that the lowland areas are completely unsuitable, then yes, you should probably exclude those areas from your analysis. However, you don't want to confuse "unsuitable habitat" with "unused but potentially suitable habitat". If there is a chance that goats might use some of those areas, then you should probably keep those areas in the analysis. Or perhaps, you could run both sets of models (with and without lowland areas) and discuss the results from each set of models.

Yes, just model mountainous areas

Thank you all.

:)

Dear Brian,

Thanks a lot. The question of my study is that were can goats appear in the park, but I didn't expect this result. Let me explain a little bit. Water sources in the area is truly very important and as Andrea said I think those places can be suitable, considering just the variables I used, meanwhile they are vacant. I used three methods: Random Forest, Classification and Regression Tree, and MaxEent. In all of the methods, elevation and water were compatibly the first important variables. But the thing is in my opinion BOTH factors should exist for the species, because. the species will be hunted as soon as it comes to lowlands. Height means safety for it. And I think if I could add a "game-species-distribution" variable, it wouldn't show any part of lowland as a suitable area.

Do you think maybe I could give a higher weight to elevation or slope (taking into account that I couldn't sample high slopes while the slope is important in goat safety) in the process of modeling and still model the whole park? Does that solve anything?

Hi Shabnam,

I agree with Brian, but given your last answer about the modelling methods you adopted, I suspect you used presence-only data with background points (BP). If this is the case, extent of study area, sampling bias in presence records and BP selection are key issues that should be accounted for, in order to avoid biased predictions, such as high suitability of lowlands areas. Thus, if you have biased predictions, removing lowland areas will not remove bias from mountain areas neither a biased estimate of variables importance .

Conversely, if you used presence-absence data, MexEnt is not a proper method (see Guillera-Arroita et al., 2014, MEE).

Moreover, did you validated your models with independent data or cross-validation techniques? How was the predictive power of your models?

You can find a lot of literature about that, but if you need I can give you some suggestion to start from.

Valerio

Thanks dear Valerio,

I used presence- only data to model and the method of testing the model was cross-validation.

I appreciate your suggestion. :))

Hello Shabnam,

My last Masters student conducted a landscape ecology study of crowned lemur (Eulemur coronatus) habitat use in a small protected area (Oronjia Conservation Area) in northern Madagascar. He modelled the habitat use of the Oronjia crowned lemur population in MaxEnt, utilizing both presence *and* absence data that he collected by walking census transects across Oronjia forest and recording where he encountered crowned lemurs (and what the forest sttructure was like), as well as keeping track of those areas where he *never* saw the lemurs. The literature on habitat use modelling with MaxEnt adopts the position that the models generated will be stronger if *both* presence and absence data are incorporated into the models. If you know the general altitudinal zone above which the wild goats are generally found, you could designate habitat below that altitude as "not used" (i.e., the wild goats are absent). Interestingly, in my Masters student's study, proximity to water sources was also a key environmental variable in regards to habitat use -- even though crowned lemurs are primarily arboreal!

Thank you dear Ian.

Hi again, Shabnam.

Although I’m not experienced with MaxEnt, Ian’s comments make good sense to me. I would expect presence-absence data to be more instructive than just presence data (my initial comments were geared towards presence-absence data). Specifically, if the data include absences/vacancies in the lowlands, the models will presumably and rightfully “discount” the suitability of those areas. Also, going back to Valerio’s question, cross-validation might be a good way of doing “triage” in the event model predictions still defy expectation. For example, if the models continue to predict high suitability in the lowlands, and have high predictive accuracy, you might look towards the data (rather than the model) for an underlying source of that apparent bias.

Brian

Dear Shabnam,

I agree with Ian and Brian that the use of presence-absence data gives better models, but I would like to clarify the issue on the use of MaxEnt with presence-absence data, as it could be helpful for both you and other readers. The debate between Thibaud et al. (2014, MEE) and Guillera-Arroita et al. (2014, MEE) demonstrated that the raw output of MaxEnt is a relative suitability and not a probability of occurrence, even If you use it with “true” absences as background points. You need post-processing of MaxEnt raw output to convert it into a probability of occurrence. I did not find the literature mentioned by Ian, but I cannot exclude that different conclusions could have been drawn from a different study framework, as MaxEnt is extremely sensible to available data and model settings.

In general, I think that modelling presence-absence data with a standard regression analysis could be easier and more directly interpretable.

Anyway, given that you used presence-background data, I guess you do not have true absences, and these comments could be a take-home message for the next project.

Valerio

Thanks you Brian and Valerio.

The model has a high predictive accuracy, above 0.9. I believe this high value is due to high overfitting, and that's the other reason I thought maybe I should limit the modeling area to the mountains. As I once told, the area is mostly consisted of lowlands. Almost 2.5% of the whole area is mountains and the results showed less than 0.1 of the whole park as suitable. See? This model is overfiited and this is the reason why the AUC rate is so high!

Another problem as Brian said may be the data. I deleted number of occurrence points because their distance were less that a kilometer. I did this to avoid overfitting, taking into account that the home range of wild goat varies so much in different places and in my study area is unknown and variables' pixel size is 30 meters. The reason I deleted data with this logic is just that I read in some articles people delete presence points closer than one kilometer to each other.

I'm pretty sure the predictive in high altitudes is correct. But, I can't deny the overfitting and the fact that some areas in lowlands are suitable. I can sort the lowlands as "not used" as Ian said. And I can't provide absence points (I try to do it in my next project :)).

So you say it's not wise to delete lowlands from modeling? What is your suggestion?

Thanks a lot. =))