Tile dft_stac_fetch to bound download over sparse-floodplain bounding boxes

[NewGraphEnvironment]

Problem

dft_stac_fetch() builds a single gdalcubes cube over the AOI's bounding box and masks to the polygon afterward. For a thin, spread-out AOI — a floodplain following a river corridor — the bounding box is largely empty, so the download scales with the bbox, not the AOI. Large-floodplain fetches are download-bound (~30 min per group at 10 m × 3 yr).

Measured on an io-lulc 10 m floodplain fetch (NECR): the grid is 57.0 M cells, of which 5.73 M (10.1%) fall inside the floodplain — a ~10× download overhead. Thinner / more diagonal reaches are worse.

Distinct from #32

#32 restores polygon-tight compute in dft_stac_cube() via gdalcubes::filter_geom(), which is blocked upstream (segfault in gc_exec_worker, confirmed on gdalcubes 0.7.4 — reported upstream). This issue is the filter_geom-independent path: cut the download by tiling, on the categorical dft_stac_fetch() (io-lulc) path.

Proposed

Tile the AOI into a grid, fetch only tiles intersecting the AOI (skip empty tiles), and mosaic. Peak download approaches the union of intersecting tile bboxes — near the AOI footprint for a corridor, rather than the full bbox. Tile size trades request count against per-request waste.

Reprex

library(drift); library(sf); library(terra)
# thin AOI, large bounding box: a ~50 km diagonal reach buffered 100 m
line <- st_sfc(st_linestring(rbind(c(-122.9, 54.0), c(-122.4, 54.4))), crs = 4326)
aoi  <- st_as_sf(st_buffer(st_transform(line, 3005), 100)) |> st_transform(4326)
r <- dft_stac_fetch(aoi, source = "io-lulc", years = 2020)[["2020"]]
nn <- terra::global(!is.na(r), "sum", na.rm = TRUE)[1, 1]
c(grid_cells = terra::ncell(r), aoi_cells = nn,
  aoi_fraction = round(nn / terra::ncell(r), 3),
  download_overhead_x = round(terra::ncell(r) / nn))
#> the fetch downloads ~1/aoi_fraction more pixels than the AOI needs

Related

• dft_stac_cube: restore AOI-polygon clip (filter_geom segfault workaround) #32 — polygon clip via filter_geom (blocked upstream)
• LULC transition/classify OOMs on large-floodplain AOIs (full-grid values + per-patch polygons) #34 — transition memory fixed; the fetch is now the dominant cost (download-bound)