2

I would like to know if we can consider as continuous data vegetal layers when disposed on seperate different .asc files (e.g.: Barren.asc; Water.asc; Evergreen Broadleaf Trees.asc; etc.)? The data on these maps seem to be in percentages (from 0 to 100) and I am working on Maxent. Depending on weither you consider the layers categorical or continuous, the output results are quite different. Also, it is mostly adviced to consider vegetation layers as categorical but I do not know if it is appropriate as the layers are in percentages and not categories. Here is a link to the layers I use: http://www.earthenv.org/landcover

Zack B.
  • 21
  • 2

1 Answers1

2

I do not know where you picked up the idea that it is "advised to condenser vegetation categorical" but, this is not empirically justified. For nonparametric models, I always tell my students to derive a multiscale proportional representation of specific classes in categorical data that they have developed hypothesis around. This, on the whole, produces much more stable models where inference can be drawn around a given nominal class. With the data represented as a set of gradient functions, high dimensional interactions can also be Elucidated, which is just not possible with a single multi-level categorical variable.

I would expect results to be notably different between an estimate derived from a continuous and categorical representation of a process. Since it is a discrete pattern, a categorical variable will smooth the variation of all the other independent variables thus, producing a conditional estimate. I do not find this a desirable characteristic in a model. Representing the process as proportional allows for a gradient function in the model and teases out considerably more in the way of functional relationships, especially when the process is represented across multiple scales (ie., focal windows sizes).

This paper may be helpful:

Cushman, S.A, K. McGarial, K. Gutzwiller and J.S. Evans. (2010). The gradient paradigm: A conceptual and analytical framework for landscape ecology. Chapter 5 in S.A. Cushman and F. Huettman (eds). Spatial Complexity, Informatics and Wildlife Conservation, Springer, New York.

Also, dig into the literature and then, based on several notable papers that show its parametric equivalence, don't use Maxent. It is, by far, not the only option.

Edit (07-17-2018 12:05): The website that you are accessing the data from clearly states:

All data layers contain unsigned 8-bit values and the valid values range from 0-100, representing the consensus prevalence in percentage.

The methods detailed in the associated manuscript present a method for consensus classification using sub-pixel classification. The values that you are seeing are not percentage of vegetation but rather, related to the percentage of class membership that specific class occupies in a given pixel. If you want a presence/absence of a given class you should just apply a threshold to each class raster. I would not however, use the raw values because in relation to species ecology they mean very little. That does not mean that you cannot then calculate the proportion of a given class within an NxN window.

Jeffrey Evans
  • 31,750
  • 2
  • 47
  • 94
  • Yes, you should apply a threshold for each veg type raster for a [0,1] outcome. The values are associated with fractional membership and not amount. I would look at a few different thresholds to see what works. Perhaps start with p>=65. Maxent creates an array of background values and is not a presence-only model, you just do not have to deal with creating the pseudo-absence data. There are much better ways to create the pseudo-absences than a gridded background. – Jeffrey Evans Jul 17 '18 at 19:52