For very large areas with sparse data, it can be computational burdensome to save this enormous file, much of which may be "thrown out" when doing the eventual training (if it is background). I'd like to add the --sparse option which mimics the operation from images to only save a handful of non-class (background) tiles