kaitlyngaynor
commented
1 year ago @amcinturff here's another thing we can discuss— do we want to look at spatial correlations between various behaviors & the habitat at HREC? this would provide a slightly different set of insights than what we are doing now (look at the relative probability of being in a given state (relative to other states) at a given covariate value)
I reran the kernel density estimation for the three behavioral states. Originally, I'd been subsetting the points into three batches based on the most probable state of each point, but I realized that a more appropriate way to propagate model uncertainty would be to use ALL points for the KDE for each of the three state rasters, but weight each point by the model-derived probability of being in that particular state. The end product looks fairly similar, so we're okay, I think.
But I wanted to check in with you @amcinturff on a couple of decisions to be made:
1 - what should our kernel bandwidth be? I ended up using 400 meters, because that's what we had previously determined (for the JAE paper) to be the maximum distance at which a hunter could hunt a deer, so feels biologically meaningful (while also capturing GPS error). But I'm not sure I'm totally sure what a 'bandwidth' is.
2 -what raster resolution should we use for calculating these rasters, and then examining their correlation with environmental covariates? The environmental covariates are all at 10m resolution, but that feels wayyy small for this kind of analysis. 50m? 100m? Not sure how to decide.
3 - there's the question of how to calculate the correlation between KDE and other rasters. I was thinking of using
spatialEco::raster.modifed.ttest()(see documentation) --- is this the method that you used, @amcinturff ? And what 'neighborhood' does it make sense to account for? Could go withd == "AUTO"but I'm not clear on what that does, exactly