Apply a fitted MoCo v2 encoder over a full raster mosaic — foundation_moco_embed

Slides a patch_size x patch_size window over the input raster at stride stride, extracts each patch, encodes it with the backbone of moco$encoder_q, and writes the resulting embedding vector back as a multi-layer terra::SpatRaster whose layer count equals moco$feature_dim. Patches whose centre cell is NA in the input produce an NA row in the output.

Usage

foundation_moco_embed_raster(
  moco,
  stack,
  dataset,
  patch_size = NULL,
  stride = NULL,
  projection = FALSE
)

Arguments

moco: An edaphos_foundation_moco returned by foundation_moco_pretrain_tiles().
stack: A terra::SpatRaster with the same channel order as the one used at training time.
dataset: The edaphos_tile_dataset used at training time (provides the per-channel normalisation statistics).
patch_size: Integer – must equal dataset$patch_size.
stride: Integer step size of the sliding window (default half of patch_size).
projection: Logical – if TRUE return the L2-normalised projection-head outputs instead of the backbone features.

Value

A terra::SpatRaster with feature_dim (or proj_dim) layers.

Details

Normalisation uses the global per-layer mean and sd stored in dataset, so the embedding pipeline is consistent between pretraining and inference.