Minimum training presence area meaning, is it a percentage of total area?

[Fabio Berzaghi]

I have a very simple question but I cannot seem to find an answer

. Does the "Minimum training presence area" value represent the percentage of the total area?

Because in my case things do not match up when I apply the minimum training presence threshold to create a binary map.

For example

, my study area is a raster of roughly 4.7 million cells. The Minimum training presence area for my model is 0.0372, which should be close to 4% of the total area, roughly 176,822 cells. Now when I create the binary map I only get about 31,000 cells that represent the minimum training presence area. This is a huge discrepancy. I have applied the reclassification procedure correctly in ArcMap several times.

Can someone please shed some light

?

[David Galbraith]

Not sure, I've struggled quite a bit with unexpected results when using the reclassify tool. Could it be a precision issue behind the scene with the exact value of minimum training presence and which side of the threshold the split is made?

--
You received this message because you are subscribed to the Google Groups "Maxent" group.
To view this discussion on the web visit https://groups.google.com/d/msg/maxent/-/A9UyAHWnMlwJ.
To post to this group, send email to max...@googlegroups.com.
To unsubscribe from this group, send email to maxent+un...@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/maxent?hl=en.

[Fabio Berzaghi]

The problem is not the reclassify. I have tested this by just converting the ASCII into a raster in arcmap. Then I used the symbology and created two ranges and I get the same output as the reclassify.

 

At this point maybe the meaning of the maxent value is unclear. Is it just refering to a percentage of the training points?

[David Galbraith]

I think the 'minimum training presence area' rule refers to a threshold value (logistic, not raw) identified by the lowest habitat suitability estimate output in your training data set. Area would probably be calculated on your background study area, training points/cells inclusive, but this calculation can be a bit complicated if you are working in a geographic coordinate system (the longitudinal distance, or x component of area, is smaller as you travel away from the equator).

To view this discussion on the web visit https://groups.google.com/d/msg/maxent/-/TZcFgK7EHNoJ.

[Matthew Aiello-Lammens]

As far as I know, David is correct concerning the fact that 'minimum training presence' is rule used estimate a thresholding value.  I'm not sure how exactly Maxent determines the 'Fractional predicted area', but if you threshold the resulting suitability map, making it binary, the number of cells in the predicted area should be about the same as the fractional predicted area times the total number of cells, with one caveat.  You have to account for the fact that some of your cells have NA values, especially if you are working with a raster layer that includes areas in the ocean.  Out of curiosity I tested this on a recent model I ran.  I used the raster package in R do the calculations, and here are some results.

Min. Train Pres threshold = 0.007

Fractional predicted area = 0.706

Total number of pixels in my map = 3456000

Without accounting for NAs, number of pixels I would expect to be classified as present: 0.706*3456000 = 2439936

Actual number of pixels greater than 0.007 = 1698767

Total number of NAs in my map = 1068449

Accounting for NAs, number of pixels I would expect to be classified as present: 0.706 * (3456000-1068449) = 1685611

These numbers are off by 13156, which is less than 1% of all pixels that aren't classified as NA.

So Fabio, my suggestion to you is to see if you have any NAs in your raster (of some other value that is a placeholder for NODATA)

---------------------------------------------------------------
Matthew Aiello-Lammens
Graduate Student in Ecology and Evolution
SUNY at Stony Brook
mlam...@life.bio.sunysb.edu
matt.l...@gmail.com

[Fabio Berzaghi]

Thanks Matt for your thorough reply. Where do you see the 'fractional predicted area' value? Are you referring to any of the area values?

I use 30-arc second data, so it's very easy to compute area since every cell is one square km. I also use the .avg of the model, but that doesn't really matter. Here is a breakdown.

 

total number of cells = 4,803,840

total number of cells minus NA values = 4,753,286

 

min. training presence area = 0.0372

min. presence logistic threshold = 0.05

 

expected number of cells above threshold = 4753286 * .0372 = 176,822

 

actual pixels in map above threshold = 31,000

 

There is a huge discrepancy whether you use the total with or without NA values. With another model of a different species the discrepancy is smaller but still by too much. expected is 43,254 and actual is 31150. Again there is something wrong here, maybe I am missing something.

[Matthew Aiello-Lammens]

Hi Fabio,

I'm not sure why you are seeing such a large discrepancy - that seems odd to me.  I got my 'Fractional Predicted Area' from the table of different threshold values in the html summary file.  Applied the logistic threshold value to the *.asc file that has the resulting suitability values, using the logistic output.  So in my case, the file is simply called Frangula_alnus.asc.  It seems unlikely, but did you change the Output format to something other than 'Logistic'?  That may change what the values are in the main *.asc file.

---------------------------------------------------------------
Matthew Aiello-Lammens
Graduate Student in Ecology and Evolution
SUNY at Stony Brook
mlam...@life.bio.sunysb.edu
matt.l...@gmail.com

To view this discussion on the web visit https://groups.google.com/d/msg/maxent/-/2wiZFQIPRMsJ.