I am building species distribution models using machine learning models based on GBIF data (presence-only data) and working on a very large spatial scale, encompassing all of North America. Before constructing the SDMs, I applied environmental filtering based on bins created by splitting continuous environmental variables in the dataset. The filtering process aims to reduce potential impacts of spatial sampling bias and decrease spatial autocorrelation in the data.
In constructing the SDMs, several methods of spatial resampling are suggested to partition the data and accommodate autocorrelation. Given that environmental filtering already mitigates autocorrelation, I wonder if the use of spatial resampling is still relevant. Therefore, several scenarios come to mind, but I'm unsure which is best:
- Apply environmental filtering and use a simple random resampling method.
- Skip environmental filtering and use a spatial resampling method. For instance, considering spatial blocking (e.g., rectangles, spatial polygons, and buffers) which might account more for spatial autocorrelation than environmental blocking (if I am not mistaken).
- Apply environmental filtering and use a spatial resampling method (potentially environmental blocking).
I tend to favor option 3, but I would greatly appreciate your advice. I pose this question because environmental filtering reduces the size of the dataset, which could pose issues with cross-validation (for some models, I receive warnings about insufficient observations).
