Hi Robert, I've produced lots of relatively poor-performing distribution models using Maxent and I think what is happening in many cases is biological and methodological. Methodological because your explanatory or predictive variables aren't the right ones, or don't capture distributional constraints adequately. Biological because the species of interest is a generalist in some niche dimensions, and the model is indicating that the predictor or sets of predictors is relatively uninformative because biologically the range that they exhibit doesn't impact distribution. Bit convoluted but hope that makes sense.

Hi Robert, I've produced lots of relatively poor-performing distribution models using Maxent and I think what is happening in many cases is biological and methodological. Methodological because your explanatory or predictive variables aren't the right ones, or don't capture distributional constraints adequately. Biological because the species of interest is a generalist in some niche dimensions, and the model is indicating that the predictor or sets of predictors is relatively uninformative because biologically the range that they exhibit doesn't impact distribution. Bit convoluted but hope that makes sense.

First, I noticed a rather unusual Regularization Multiplier (5) instead of the default (1). So far I have only used the default or 2. I assume you have tried the default value? A methodological issue (see Paul's answer).

Next, I wonder how large and divers your modeling area is. If rather small and/or homogeneous, the usual macro-bioclimatic variables and major vegetation cover categories, for example are less likely to be good predictors.

Finally, I had more than once to assume (for plants) on the micro-scale that anthropogenic legacy effects (for which proxy variables are often hard to conceptualize and estimate ) would perform better than the usual set of biophysical predictors.

Thanks Paul and Hein for your replies. I used a multiplier of 0.5 (typo) and 1. The species has been reported in the literature from a wide area of Amazon forest and Atlantic forest (while there are no observations from the band of xeric vegetation between both), so the area is large and divers. I have collected this species in several locations of moister forests in a transition to Amazon forest while it appears to be absent in drier forests in my study area - no observations. Based on this I assumed that the usual macro-bioclimatic variables are good predictors. However, it appears to be that there is something else, biological or methodological. Might be more generalist than a highly host-specific herbivore...

The only solution is to revise or make another model to increase the AUC. And improve parameterisation

Hi Roger, my suggestion is that you can improve your model based on a multi-step process (see paper Ramirez et al., 2018). In addition, you can use the approach developed to low presence based on paper reported by Pearson et al., (2007). Another factor that can affect the performance of the model is the resolution of environmental variables.

Ramírez-Gil J.G., Morales J.G., Peterson A.T. 2018. Potential geography and productivity of “Hass” avocado crops in Colombia estimated by ecological niche modeling. Scientia Horticulturae 237:287–295. doi: 10.1016/j.scienta.2018.04.021.

Pearson RG., Raxworthy CJ., Nakamura M., Townsend Peterson A. 2007. Predicting species distributions from small numbers of occurrence records: a test case using cryptic geckos in Madagascar. Journal of Biogeography 34:102–117. doi: 10.1111/j.1365-2699.2006.01594.x

Dear Robert, a part from all methodological issues that I believe you already considered, especially about the AUC but even about the Maxent itself (pseudoabsences etc.), I always believe that best answers can only come from people that knows the target species. If you know the ecology, threats etc of your species so you are the best for explaining (and/or ameliorating) such results.

Of course it is not the only paper about the use of Maxent for data-poor species, but I would like to share with you my small contribution.

Regards

Article Using species distribution models at local scale to guide th...

Hi Robert,

I have developed SDMs for several species and found the extent of study area matters. If your study area is small, the AUC value is low. One solution is to develop an ensemble model: combine predictions of several single modelling techniques. You can construct an ensemble model by biomod2 R package.

regards,

Zhixin Zhang