HELP ON REDUCING SPATIAL AUTOCORRELATION

321 views
Skip to first unread message

gafna jeff

unread,
Sep 25, 2019, 9:52:11 AM9/25/19
to Maxent

 Hello everybody,

I am modeling some plants with huge occurrence records (as many as 5,000) using maxent in R program. Due to sampling bias, spatial autocorrelation (SAC) in residuals of my maxent model are clearly evident (shown by the Moran’s I). It seems that the SAC is largely as a result of SAC in the variables that I used (bioclim and soil) and not SAC in the occurrence records. Despite using R packages that thin or reduce occurrence records in the best way possible (so as to reduce spatial autocorrelation) such as spThin, thin.max and spatialEco, I have been unable to the reduce spatial autocorrelation. I also used SAM to select spatial filters that were added to my model as a raster but this did not help. Could anyone suggest a function or a package that can help in reducing SAC in residuals of a maxent model. How could, I go about this.


Regards,

Jeff  

Alicia Langton

unread,
Sep 29, 2019, 5:26:49 PM9/29/19
to max...@googlegroups.com
Hi Jeff,

I was thinking about your SAC again and see that you did not mention creating a bias file for your data. The bias file prevents the randomly generated background samples from being created outside the regions of your area that were sampled. This increases model accuracy, but may not directly address your SAC.  If you think this could be useful, I will send you my R code.

Best,

Alicia

--
You received this message because you are subscribed to the Google Groups "Maxent" group.
To unsubscribe from this group and stop receiving emails from it, send an email to maxent+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/maxent/28349f58-e0ee-4111-8f75-94023540ece3%40googlegroups.com.


--

Jamie M. Kass

unread,
Sep 29, 2019, 5:35:38 PM9/29/19
to Maxent
Yes, look into sample bias rasters. But I think you are misunderstanding SAC. If points are close in space, they will usually be correlated when considering a spatial predictor variable. I'm not sure what you meant when you said your points are not correlated but your variables are. The correlation is measured with the variables. You should see drops in SAC when you do spatial thinning or run things like SAM. Was this not the case?

Jamie

gafna jeff

unread,
Sep 30, 2019, 6:12:21 AM9/30/19
to max...@googlegroups.com
Dear Alicia,
Many thanks for your time and email. Sure, I have never considered creating bias file for my data. Probably this could be a viable option. Just wondering if selecting my background points outside the regions that I sampled (as you suggest) is any different from creating a buffer zone when selecting my background points. The buffer zone could even be 10 km from each presence points. I would really appreciate your R code.

gafna jeff

unread,
Sep 30, 2019, 6:23:55 AM9/30/19
to max...@googlegroups.com
Dear Jamie,
Many thanks for your email. We just agreed with my supervisor that, Its probably impossible to reduce SAC in residuals of a maxent model. But he has suggested that I should reduce SAC in the occurrence records. From my understanding, we have two types of SAC - SAC in occurrence records and SAC in variables used. To deal with SAC in variables, one must reduce SAC in residuals of the maxent model (this has been impossible). I conducted spatial thinning of my occurrence records using spThin and thin.max packages but still had SAC in residuals of my model. For now,  I would like to check if SAC is still there in my occurrence records after thinning but am struggling with writing the R code of conducting a Morans test on the remaining occurrence records (after thinning). Do you have some code in this regard?. I would really appreciate.

Regards,
Jeff


On Sun, Sep 29, 2019 at 11:35 PM Jamie M. Kass <ndimhy...@gmail.com> wrote:
Yes, look into sample bias rasters. But I think you are misunderstanding SAC. If points are close in space, they will usually be correlated when considering a spatial predictor variable. I'm not sure what you meant when you said your points are not correlated but your variables are. The correlation is measured with the variables. You should see drops in SAC when you do spatial thinning or run things like SAM. Was this not the case?

Jamie

--
You received this message because you are subscribed to the Google Groups "Maxent" group.
To unsubscribe from this group and stop receiving emails from it, send an email to maxent+un...@googlegroups.com.

Surajit Hazra

unread,
Sep 30, 2019, 11:24:18 AM9/30/19
to max...@googlegroups.com
Dear Alicia Can you send me your r code?

Thanks in advanced

Areej

unread,
Feb 10, 2022, 6:09:41 AM2/10/22
to Maxent
Dear Jeff

Can you explain for me how you estimated the residuals from your model? As I'm trying to estimate it but I'm not sure if my steps are correct

My steps:: 
 The first thing is that the model prediction values comes in ascii format. So I was thinking on how I'm going to substract that from the observed values (i.e cells with value of 1)? 

Since I'm using ArcGIS, my initial trials were to create a raster layer with only presence cells that are equal to 1. The produced layer appeared as separate cells in empty space (as I read the residuals are only calculated for presence cells only). Then I tried subtracting maxent ascii layer (which contains the predicted values) from this layer. The obtained difference layer is then converted to a point/polygon feature format so I can conduct Moran's test (Moran's test accepts features only). 

Are these steps correct?  and if you are not familiar with ArcGIS can you explain your alternative method even if it was manual?

Best
Areej

Reply all
Reply to author
Forward
0 new messages