Coarsen a raster or dataset by block means.
coarsen.RdMaps x onto the grid factor times coarser, anchored on its origin:
each coarse cell is the mean of the valid (non-NaN) values of its
factor x factor block, NaN when none is valid. This is exact block
arithmetic, not a GDAL resampler, so the two stages stay explicit: a
decoded embedding read by nearest neighbour onto a 10 m grid, then
averaged to 30 m,
Details
codes <- lazy_dataset(tiles, grid10, resampling = "near", decode = dequantize_aef)
cube <- stack_bands(coarsen(codes, 3L))takes, for each 30 m cell, the mean of the nine decoded 10 m values the
10 m read gives, never a GDAL average over the reprojected source
footprint. The coarsening runs inside the read: each window's finer
cells are assembled from the sources as the nearest-neighbour read of
them assembles them (a direct window read of a source on those cells'
pixels, a nearest warp of any other source alone, later sources
winning where they hold data), decoded, and averaged in one pass, so
the finer grid is never written. Cells warped from a reprojected
source match the finer read's under garry.warp_exact = TRUE; with
GDAL's approximate transformer (the default) a cell near a pixel
boundary can take the neighbouring pixel.
Version 1 coarsens reads only: x must come straight from
lazy_dataset() or lazy_source() (decoded or not) by
nearest-neighbour, from sources opened without open options, and its
dims must divide by factor. A dataset with a mask_asset is
refused: averaging a mask band's bits would invent flags.