You can find out the weight of each variable in the prediction from the response curves. Previusly, set "create response curves"

References: 

Phillips, S. J., & Dudı, M. (2008). Modeling of species distributions with Maxent : new extensions and a comprehensive evaluation. Ecography, 31, 161–175.

Elith, J., Phillips, S. J., Hastie, T., Dudík, M., Chee, Y. E., & Yates, C. J. (2011). A statistical explanation of MaxEnt for ecologists. Diversity and Distributions, 17, 43–57. http://doi.org/10.1111/j.1472-4642.2010.00725.x

Merow, C., Smith, M. J., & Silander, J. a. (2013). A practical guide to MaxEnt for modeling species’ distributions: what it does, and why inputs and settings matter. Ecography, 36(10), 1058–1069.

Radosavljevic, A., & Anderson, R. P. (2014). Making better Maxent models of species distributions: Complexity, overfitting and evaluation. Journal of Biogeography, 41(4), 629–643. http://doi.org/10.1111/jbi.12227

Thanks Oliver. I've attached the .png of the response curve for this land use/cover variable. Clearly each class in process individually. But then what does MaxEnt do with the lulc layer -- I am starting to assume it is handled as an aggregate, instead of singling out each land class. Therefore is it correct that the presence points that are on one of the high-frequency classes increase the influence of the lulc layer (as a whole)? Or how does presence relate to category?

Maxent does not asume all the classes of the land use layer as an agregate. For example, you can see that there are some Land uses classes (21,41,90) more important that others in your prediction for Linaria dalmatica. This is the meaning of response curves. However, Maxent prediction uses "backgrounds" data. These background-data are used by Maxent like an "agregate" that depends of extent (study area). Anyway, the question is still open because it is not clear to me the meaning of "agregate". I hope I have been able to help you

Hi, Oliver, by aggregate I mean this: the response curve (my .png) is calculated for each class; however, MaxEnt processes all the binary features created for the categorical variable and reports the results as one result. So e.g. the jackknifes just show the land use/cover categorical variable as one (instead of something for each class). This means, to me, that MaxEnt reports results for the categorical variable as one variable - in the same way other variables/features are reported. Therefore the influence of the categorical variable is measured the same way the influence of a continuous variable is measured.

Is that correct??

Now I understand. Yes. You can adjust the distribution by changing Setting / regularization multiplier. For example, a smaller value than the default of 1.0 will result in a more localized output (overfitting?). On the contrary, a higher value than the default...will give a more spread out distribution. This is very related with categorical variables. 

You could create new dummies variables (layers) from the most important classes of land use. Another option is to realize a transformation of categorical variables, for example "evidence likelihood"

Finally, you could realize an experiment with two virtual species and compare the results regarding to the influence of land-use layer. Two ways. 1. Virtual species based on random points 2. Virtual species based on localized points.

Thanks Oliver; I appreciate your help!

Indeed, the jackknife test shows a categorical variable as a single variable at par with a single continuous variable, irrespective of the number of categories/classes/types.

Presumably, the response "curve" (diagram) shows, in analogy with the curve of a continuous variable, the probability of occurrence of the species within each of the 17 lulc classes in the example.

Only the cluster of four lulc units after 21 on the horizontal axis and the last of the cluster of lulc units after 41 show a high probability.

Unexplained remains the meaning of the numbers along the x-axis from zero (0) to presumably 100. More so as we have encountered outputs with a value >100 along this axis in our work. Further, the apparent clustering of the lulc units in the diagram seems unexplained.

Hi,

I'm going through something similar. I ran Maxent using categorical variables that I extracted from an aggregated dataset (SOLRIS). So I had wetlands, beach habitat and tilled fields as binary categorical data (1 = wetalnd, 0 = no wetland) because I didn't want everything in the dataset, which had 50 different habitat classes. And I put these 3 layers into Maxent. I also put a continuous variable in the model, a DEM showing terrain ruggedness. However this caused problems. Maxent models ended up running on the continuous variable and barely acknowledged the categorical variables. I'm assuming this is because of the binary nature of the 3 habitat layers.

I recognize Maxent will NOT run on binary categorical data. How do I change this so Maxent will run each categorical variable AND the DEM and it won't take priority of the DEM.

Below are maps of all my environmental variables, where the patches of colour show the presence of the habitat and anything that's not the habitat will be given a cell value as "zero", so you see the topography map of Southern Ontario.

Thanks for the wonderful and detailed replies.

Quantitative environmental factors from the categorical variables is a suitable method to input the Maxent to build the SDM.namely quantifying the binary categorical factors via the regularization multiplier and creating a new layer then containing the need variables.

Rebekah Persad What's the tips for the error that not running during the Maxent?

That could be the different coordination and spatial extent resulting in the problem.If so, try to resample the environmental variable layers in the ArcGIS.

HTH,

Chen

Chen Xiao Fish

Hi Chen,

I am still clueless on how I would have done it differently, but I ran out of time to complete a better and more accurate analysis for my project.

Essentially I was told from an expert that Maxent CAN run on binary variables, I think it's a matter of laying out the variables on a grid using R (programming language) or formatting it in a different way so that the "No Data" cells would be disregarded and all the environmental variable cells should be treated equally. So all the variables would be treated as individual cells and the background cells are ignored.

I still ran the analysis without doing this and it produced three maps. As you can see, it's messy. Not the best results, but I was able to look at areas of high suitability vs lower suitability.

I did make sure all my layers had the same spatial extent and coordinate system. I also resampled all the layers, and resampling actually messed up the habitat features in my environmental variables. It made things choppy. Instead of having a delineated beach shore, resampling made the beach shoreline into a large blocky cell (see second image below). In this image, the original shoreline are the white cells, and the resampled shoreline is the green cell.

I'm finished with the project, but it'd be nice to know and see where I went wrong.