Build a lazy patch dataset over a terra::SpatRaster — foundation_tile

Returns an edaphos_tile_dataset S3 object that knows how to sample batched (B, C, H, W) tensors from the underlying raster on demand – without ever loading the whole mosaic into memory. The dataset is what foundation_moco_pretrain_tiles() consumes.

Usage

foundation_tile_dataset(
  stack,
  patch_size = 16L,
  n_patches = 1000L,
  valid_mask = NULL,
  normalise = TRUE,
  seed = NULL
)

Arguments

stack: A terra::SpatRaster with the channels to sample.
patch_size: Integer – spatial side of each patch.
n_patches: Integer – total number of patches the dataset should expose. Sampling is with replacement when n_patches > total valid cells.
valid_mask: Optional terra::SpatRaster of 0/1 defining which cells are eligible as patch centres.
normalise: Logical – if TRUE (default) each channel is standardised to zero mean / unit standard deviation based on the raster's global statistics.
seed: Optional integer for reproducibility.

Value

A list of class edaphos_tile_dataset with slots stack, patch_size, n_patches, n_channels, means, sds, valid_cells; and a sample(batch_size) function that returns a (batch_size, C, H, W) R array.

Details

Patch centres are drawn uniformly within the AoI, optionally constrained to a valid_mask (e.g. a land-sea mask or a biome polygon). Each patch is a patch_size x patch_size window extracted from the raster; missing values are replaced by the per-layer mean.