danddr / use Goto Github PK

Fitting a PCA on large rasters can take a lot of time, so it could be useful to allow for a subsample of data to be used. This was done in RSToolbox via the nsamples argument. I believe this feature would only require a minor change to the code here, by allowing for a user specified value in rastPCA to be passed to the maxcell argument of terra::layerCor.

I noticed a minor mistake and something that could be made more efficient in rastPCA.

In rastPCA, an NA mask is created using maskNA <- is.na(env.rast[[1]]), but the [[1]] indicates that it is only checking for NA values in the first layer rather than all layers. RSToolbox uses totalMask <- !sum(calc(img, is.na)) to check across all layers. RSToolbox uses the raster package, but the equivalent terra code would be totalMask <- !sum(app(img, is.na)).

I also noticed that the entire raster is coerced to a dataframe to get the sample size in eigenDecomp$n.obs <- nrow(as.data.frame(env.rast[[1]])). Converting the dataset to a dataframe slows things down and isn't memory safe. It'd be better to count up the valid cells (i.e. cells that don't have NA values in any layer). In RSToolbox this is done with model$n.obs <- cellStats(!any(is.na(img)), sum). The equivalent terra code would be model$n.obs <- global(!any(is.na(img)), sum)$sum