Gurnard Parish Projected SDM

Introduction

One of the useful features incorporated into MaxEnt is the ability to form a "projection" of the likely habitat use using the parameters calculated within a modelled and tested area. There are many examples on the web for what can be achieved using the projection capability and the pitfalls that need to be avoided. Projections allow you predict likely species use of similar habitats to the one where you have sample data, and are particularly useful as a pre-curser to planning survey work or assessing areas where access is limited. I thought that I'd include an example to demonstrate this feature, so if you are interested, read on...

Setting up a projection model

Firstly, the area you intend to project the species distribution model to needs to be a similar habitat to the one you've used to generate the model parameters. It is pretty pointless trying to project a species distribution into an area with only marginal similarity. I'll say a bit more later about how you can assess the habitat correlation from the projection model, but suffice to say if the habitats have only marginal correlation, then a projected SDM is likely to be inaccurate and of little scientific value.

Assuming you already have environment grids and samples set up, then all you need in addition to these is a duplicate set of environment grids for the area you want to project to. In the model I set up for the Gurnard Parish area (located adjacent to the Parkhurst Forest) I set up a directory structure as shown below:

So, you see I've set up three folders, one for the environment files that will be used to generate the model, one for the projection area environment files, and one for the model output. The presence samples used in the model are in .csv format and placed at the top level of the model directory.

The environment files themselves need to represent the same variables, so if your land use raster has oak woodland represented by a pixel value of "4" in the model area, then the projection raster grid for land use must also use a pixel value of "4" for oak woodland. The environment grids do not have to be the same size or extent, but the grid should ideally be the same pitch. The picture below shows the structure of the files in the two environment grid folders:

You can clearly see that the file structure is the same in both the environment grid folder (for generating the model) and the projection folder (for generating the projection).

That's it basically, and if you get this far you are ready to run your projection model!

Running the model

You need to do some extra setting up in MaxEnt prior to running the model. Aside from the additional folders you need to reference, make sure that you check the write clamp grid and do MESS (MultivariatE Similarity Surface) analysis boxes on the settings page (you will need these in order to understand how well the projection worked.

The image below shows how I set the model to run for my projection into the Gurnard Parish area.

So, that's it; model ready to run!

Interpreting the results

The first thing to do assuming the model runs and completes is to look through the HTML pages with the model summary information. You'll notice is that there are additional charts, firstly those with the model projection in, and then three further charts that show how effective the projection is. The first of these additional charts will look something like this:

This chart indicates just how well (or otherwise...) the model has performed by showing in the warm colours where model data has been clamped in the simulation. In this case everything in the data area is blue, so looks pretty good. The next chart is the MESS Chart, which in my case is shown below:

The red areas indicate that there two or more variables outside their training range, so results in this area are effectively an extrapolation and therefore less certain. The main problem area in this case corresponds to a salt flood plain area around the river Luck which is at sea level and outside limits of my altitude environment in the Parkhurst Forest model. The flood plain is also outside the landuse variables I had, so the projection is not likely to be accurate in this area.

The final chart is the one that shows the most dissimilar variable or MoD map. Mine looked like this:

This clearly highlights the flood plain area again and pins the largest error to the altitude grid.

Finally, here are the charts showing the projected species distribution models for the main 8 species found in the neighbouring Parkhurst Forest area: