Dated reference imagery: true-colour composites, NDWI/MNDWI, and RGB layers in dft_map_interactive (+ HLS source)

[NewGraphEnvironment]

Problem

Checking a land-cover label or a detected change needs dated reference imagery for windows we choose: the same season in different years, or early vs late season within a year. The basemaps in dft_map_interactive() (Esri, Google, and Bing in #78) are high resolution but have no date, so they can say what a place is and not when it changed.

drift already fetches Sentinel-2 L2A cubes (dft_stac_cube(), with cloud masking, a months filter and median aggregation), but only as a single-index stack (ndvi, kndvi, ndmi). Three pieces are missing to use it as a reference-imagery source. floodplains#93 needs them for a stratified accuracy assessment of IO LULC in BC floodplains.

Scope

1. True-colour (and false-colour) composites. Return a multi-band composite (e.g. B04/B03/B02, or B08/B04/B03 for vegetation) for a datetime window and months, through the same cloud mask, cache and tiling path as dft_stac_cube(). The new_graphiti post 2026-01-08-stac-ortho-mosaics did this outside drift with rstac + gdalcubes (June–July, eo:cloud_cover <= 20, median). That code is the starting point, and should not become a second copy.

2. Water and wetness indices. Add NDWI (green/NIR) and MNDWI (green/SWIR16) to dft_index_table(). ndmi is there already. These are the direct indices for wetland drying and water extent.

3. RGB layers in dft_map_interactive(). Today x is classified rasters coloured through class_table. Accept RGB composites as switchable overlays too, with labels such as "2017 Jun–Jul" and "2023 Aug–Sep", alongside the classified layers and transitions.

4. HLS as a second cube source: split off to HLS (Harmonized Landsat Sentinel-2) as a cube source: Earthdata auth, CMR-STAC query, bit-mask Fmask #82. Harmonized Landsat Sentinel-2 through NASA Earthdata: 30 m, from 2013, 2–3 day revisit, for a pre-2017 baseline and seasonal density. A live spike (2026-09-28) established:

   • CMR-STAC rejects intersects (HTTP 500) and ignores the CQL2 cloud filter;
   • reading needs a real Earthdata Login account, because earthdatalogin's default credentials now return 401;
   • S30 and L30 differ in NIR/SWIR asset names;
   • Fmask is a bit mask, which gdalcubes::image_mask(bits=) can express.

   HLS (Harmonized Landsat Sentinel-2) as a cube source: Earthdata auth, CMR-STAC query, bit-mask Fmask #82 carries the rest.

Scale (added at the plan gate, 2026-09-28)

floodplains#93 reviews sample points across whole floodplains, not a reach, and the approach is built for that:

• Per-point chips. Call dft_stac_composite() once per buffered sample point. Each chip is small and cached on its own, so adding a point fetches one chip. A single call over all the points would span the floodplain under one cache key.
• Floodplain-wide browsing. Composites are cached as COGs, and dft_map_interactive(rgb =) accepts COG URLs served by titiler, as well as local rasters. The upload to object storage is the driver's job.
• Composites are surface reflectance (float), not display bytes, so they can later serve as classifier covariates (floodplains#93, move 6).

Depends on #80: gdalcubes was archived on CRAN, so it now installs from GitHub. Pre-existing read-path defects surfaced here are #83.

Resolution (PR #86, v0.18.0, 2026-09-28)

All four acceptance items landed:

• dft_stac_composite(): true-colour and false-colour composites, cached as COGs, with a key covering bands and months.
• ndwi and mndwi.
• dft_map_interactive(rgb =), verified live with 2017 Jun–Jul and 2023 Aug–Sep over Esri/Google.
• HLS filed as HLS (Harmonized Landsat Sentinel-2) as a cube source: Earthdata auth, CMR-STAC query, bit-mask Fmask #82 after a live spike.

What differs from the text above:

• Per-point chips work; floodplain-wide composites do not yet complete at BULK size.
  • 100 BULK chips took 84 min, median 40 s each.
  • A floodplain-wide read over BULK now gets every tile (3 h 16 min). That needed two read-path fixes that also apply to dft_stac_cube(): a convex-hull query past Planetary Computer's ~1 MiB search limit, and per-tile re-signing past its ~45-min token life.
  • Clipping the merged mosaic then fails. That is Floodplain-wide tiled composite: terra::mask cannot read the merged mosaic after a 3-hour read #88, so "floodplain-wide browsing via COG" above is not yet true at that scale.
• The composite defaults to clip = FALSE, so reference imagery keeps its surroundings.
• A window straddling Sentinel-2's 2022-01-25 offset change is refused.
• A window cannot cross the new year.

Follow-ups:

• Cube read path: date-only end bound drops the last day, no #51 paging guards, Jan 2022 cover bias, miss/hit drift, temp NetCDFs #83: pre-existing cube read-path defects.
• Composite chips run serially at ~50 s each: run sample-point chips concurrently #85: concurrent chips.
• Failed gdalcubes chunk reads are silent: a partial cube passes the empty check and is cached #87: silent failed chunk reads.
• Floodplain-wide tiled composite: terra::mask cannot read the merged mosaic after a 3-hour read #88: the floodplain-wide clip.

Acceptance

• A documented call returns a cloud-masked, median true-colour composite for an AOI, a year and a set of months. It is cached, and the cache key includes the bands and months.
• dft_index_table() includes ndwi and mndwi, with tests against known band values.
• dft_map_interactive() shows at least two dated RGB composites as switchable layers over the Esri/Google basemaps.
• HLS is either delivered or filed as its own issue, with the auth path worked out (HLS (Harmonized Landsat Sentinel-2) as a cube source: Earthdata auth, CMR-STAC query, bit-mask Fmask #82; the end-to-end read awaits an Earthdata account).