Hi,
I'm currently trying to use Maxent for roadkill analysis, but one of my colleagues raised a question that using spatially correlated datasets in maximum entropy models leads to biased results and the dataset should be tested and correlated occurrences removed.
I have no clue how to test roadkill point dataset for spatial autocorrelation to get rid of correlated points. What's the easiest and the most common way to determine which points need to be removed to avoid the bias?
If you know / have any information / articles / website for me to look over, please advise on this.
Your direct advise to solve this issue is greatly appreciated.
Best,
Bryan
Thanks Lazo for your valuable comments. I have two questions on your comments.
1) Lazo, do you think we can dramatically reduce spatial autocorrelation (SAC) issue when we select 'Random Subsampling' method in MaxEnt setting? As you know, Maxent does provide an option to use 'Random Subsampling' (there are options for 'Crossvalidation' and 'bootstrap' as well) method as sampling technique (Replicated run type),
2) Your 2016 paper says, "To assess whether this could apply to our dataset, we calculated the range of the semi-variogram per environmental variable. This range was compared with the average distance between the point occurrences of P. cinnamomiin each study site and the pixel resolution."
How did you calculate the range of the semi-variogram for each environmental variable? (is this doable in ArcGIS?) As mentioned, my dataset is roadkill point data, so roadkill locations tend to be spatially autocorrelated. I'm trying to figure out how to apply your method in my dataset. It'd be greatly helpful if you can explain me how to overcome the spatial autocorrelation of my point data.
Hi Hoegun,
Are you referring to the spatial correlation among the variables used in Maxent? Or the spatial bias in occurrence points? If you are talking about the latter, maybe you may want to use only one point per grid when modeling.
Cheers!
Eric
What specific variables are you using or you intend to use in your model?
I will be clear and concise: 1. Select a few predictors. 2. Minimize correlation among pre- dictors 3. Choose the most significant predictors. 4. Separate the presence records as much as possible. 5. Split the samples by a randomized selection procedure. 6. Check the residuals of your models.
Oliver
Unfortunately your response is not very clear or concise. The first question is predictors for what and how do you select them? Unless you have a different definition of "correlation", it is not applicable to "predictors" nor are predictors "significant" or non-significant. Presumably "residuals of your models" really pertains to residuals of predicted values and "check the residuals of your models" is neither clear nor concise.
Hoehun
Presumably you have a reason for wanting to pick "one point per grid", i.e. what is your criterion? I suspect you mean one point per pixel rather than one of the vertices of the grid. Are you going to use the value at the selected point to "represent" the value for the pixel, for what purpose? For very small pixels, using the center point might be reasonable but for a large pixel using only point might give very mis-leading results, i.e. if the values at points within the pixel vary significantly the the results from the intended use may vary significantly.
You need to think about what you want to do and not just search for a software package that seems to do it
Donald is right, I did not make it sufficiently clear and concise. Unfortunately my response was not comprehensive because of I was presupposing that Hoehun has some background and familiarity about the methodology. For that reason, I have simplified a lot. I shall endeavour to explain myself clearly.
On the first point, it is very important to justify the choice of predictors, and interpret the results in accordance with used predictors. On the second point, I am talking about multicollinearity. That is, high correlations between several predictors. On the third point, I would emphasise the importance of acknowledging the ecological significance -in statistical terms- of the selected variables in relation to the response variable. On the sixth point, I recommend plotting (histogram, qqplot) the residuals and test for normality. Another option is to study the spatial structure using the correlogram ot the residuals. The error should not have any spatial pattern. The correlogram of the residuals should be flat.
I also understand that "one point per grid" means "one point (species occurrence record) per pixel". At least that's what I think. This option is interesting to avoid pseudoreplication.
Entropy: A concept that is not a physical quantity
https://www.researchgate.net/publication/230554936_Entropy_A_concept_that_is_not_a_physical_quantity
- Badges
- Science topic