building a bias grid

[Natalia Quinteros]

Hello, 

I'm trying to build a bias raster file, but I had no luck finding how to do it, I'm working with a single species and they are samples of convenience. What I have are gps points for each individual collected and the environmental grids that I'm planing to use in MaxEnt. 

The literature says: 

"If you believe that your species occurrence data constitute a biased sample, and you have a good understanding of the spatial pattern of sample collection effort that produced your occurrence data, you can provide Maxent with a “bias grid” which is then used to correct for the bias.  The bias grid should have the same dimensions, cell size and projection as the environmental variables, and should be positive (or no-data) everywhere.  The values should indicate relative sampling effort, so if two cells have values 1 and 2, that means the probability of having visited the second cell is twice as high as the first.  Note that the bias grid gives a priori relative sampling probabilities; it does not indicate where sampling actually happened." 

then, I've been thinking that maybe there is a possibility of calculating this raster by using a geostatistical tool in ArcGIS

I would like to know how people have solved this problem. 

I'll appreciate if you could share your methods with me. 

Also, I read in the group this posted in 2007

If you have a set of samples that you think is representative of
sampling bias, such as presence data for a group of species that are
all detected using the same methods or by the same researchers, then
you can use them in Maxent by providing them as background data.  You
need to have those samples in SWD format, with the environmental data
given for each point.  You then use that file instead of the
environmental layer grids; you can project the resulting model onto
your original layers to get a prediction for the whole study area.
(We have a paper in preparation that discusses how this can counteract
sampling bias.)

If you instead use a "sampling bias" file, it needs to be a grid,
typically in .asc format.  A bias file describes the relative sampling
intensity across the whole study area.

I don't understand this part "you can project the resulting model onto

your original layers to get a prediction for the whole study area"

how do you do this projection?

Also, for the bias file (again) how do you calculate this sampling intensity? % of samples for sq km? 

Thanks for your help again! 

Natalia

[mes...@gmail.com]

Hi Natalia,

There still seems to be some mystery surrounding the method of creating a bias grid. The only actual instructions that I have found for creating it are found here and they are a little dated. For example, they tell you to utilize Hawth's tools (they use R to run) but Hawth's tools is now called Geospatial Modelling Environment (GME). I have been working through creating a bias grid myself using those instructions, but unfortunately my dataset is too large and the tool needed in GME  (distance among points) runs out of memory. I have been instructed to complete that step using just straight R. I hope this helps you, and at least you know you're not the only one having issues with this.

-Kristin

[Megan S]

Hi Kristin,

I am having the same issue using those dated instructions and my dataset is too large to use the point distance tool in the new Geospatial Modelling Environment. Could you provide any information on the R code you are using to create the bias grid?

Thanks for your help,

Megan

[Oliver Burdekin]

Hi Natalia,

I'm working on this at the moment for a single species across the whole of South America.  I'd like to build both a bias file and a background file.  I'm still having a think about the background file as I believe it will have to be an swd file.  For the bias file my workflow is going to be as follows:

>Using GIS software create a point density layer from sample occurrences.  Depending on the survey methods you can alter the radius of the point density tool (and if you are working from historic records split the samples into separate layers to alter the radius according to different survey methods).

>When doing this it is important to use a "snap raster" from your environmental layers to make sure all cells and extents are equal.

>Convert the output raster to .asc and then you should be good to go.

Once I've thought more about the background file I'll be re running my models with bias, then with background, then with both to see how the outputs are altered.

You may find the Merow et al (2013) paper useful (Ecography 36: A practical guide to MaxEnt for modeling species’ distributions:
what it does, and why inputs and settings matter) although it sounds like you may have already read it. 

I also find this tutorial useful: https://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=2&ved=0CDQQFjAB&url=https%3A%2F%2Fwww.cs.princeton.edu%2F~schapire%2Fmaxent%2Ftutorial%2Ftutorial.doc&ei=pglhU6WjGIfJsQSgkoBo&usg=AFQjCNGLRzLmkwCWYL3VM30dFpe3QBLSbQ&bvm=bv.65636070,d.cWc&cad=rja

[Megan S]

Hi Oliver,

Which tool are you using in ArcGIS to calculate point density? The problem with that method is that MaxEnt reads the values of the bias grid as 'relative sampling effort,' so a cell with a value of 2 is 2 times more likely to be sampled than a cell with a value of 1. Do you think the point density calculations create a grid with this type of information?

Thanks,

Megan

[Oliver Burdekin]

Hi Megan,

I'd use Spatial Analyst > Density > Point density.

I realise that the bias grid is relative sampling effort.  It's not really that it is "twice as likely to be sampled", I think it's more "it has been sampled with twice the effort than a cell with a score of 1".  When you run the point density it really depends on the cell size, spread of samples, and the total area extent involved.  In addition the cell values are graded depending on distance from the point, so values are rarely integers.  What I'm really trying to do is get MaxEnt to focus upon the areas that have been sampled.

For example, I have a number of samples discovered in Ecuador and a few in Peru.  Sampling effort has been considerably higher in Ecuador whereas in Brasil there has been non (going on presence alone).  That means Ecuadorian samples will have a higher score on the bias file whereas it will be lower in Peru and zero in Brasil.  Which is what I want.

As we are using museum specimens and literature reviews as part of our samples it is unfeasible to account for all the different methodologies and sample efforts from the many different studies conducted over the years.

I hope that helps and let me know if you have any other questions.  I think it's good to talk through methodologies like this.

Oliver

[Michelle Moeller]

Hi Natalia,

I found this tutorial explained nicely how to create a background raster in Arc. This is what I used to create mine.

Hope it helps,

Michelle

[nyasha.m...@gmail.com]

Thanks Michelle this tutorial has been a great help!