I am following the tutorial How To: Land-Use-Land-Cover Prediction for Slovenia and I would like to understand why they used a positive buffer (expanding boundary limits) in the first part of the tutorial; and why they used a negative buffer (when they eroded the parcels: ground truth data) in the second part.
It is not the first time I observe buffers being used like this.
Specifically about the tutorial provided as example:
The positive buffer along Slovenia borders is being applied possibly to avoid empty pixels in the edge of corresponding patches (i.e, pixels only touched in the corner by the polygon mask).
The erosion task is being applied over the reference map mask which contains the rasterized polygons and their classes of land-use-land-cover in Slovenia. The purpose of such task is explained in the tutorial; and it is two-folded: i) remove artifacts (a 1-pixel-size-class); ii) remove edges between adjacent classes. This was to avoid sampling from such areas, which makes sense when trying to avoid noise in the modelling process.
See more examples (not an exhaustive list) on using buffers, below:
Generating a Digital Elevation Model (while avoiding edge artifacts): Correcting merged LiDAR DEM with outliers in inner edges?
For aesthetics reasons (cartography): Using directional shading on boundaries for cartographic effect in ArcGIS Desktop?
Deleting gaps/slivers from polygons: Deleting small gaps (slivers) between polygons?
When working with window functions: PostGIS raster calculate average slope of 3x3 neighborhood around a geometry point