SingularityUS/commsight-lz

CommSight LZ — off-field landability fact tiles

Compressed raster facts about ground, built for the off-field landability layer in CommSight. One file per 1°×1° cell, named by the USGS 3DEP convention (n33w107 = the cell whose north-west corner is 33°N 107°W, so it spans lat 32–33, lon −107…−106).

⚠️ Advisory only — not for navigation

These tiles describe terrain and land cover. They are not a survey, not a database of landing sites, and not a landing recommendation. Surface condition, fences, livestock, crops in season, wires below the charted-obstacle threshold and any current obstruction are not modelled. The absence of a depicted hazard is not evidence that none is there. Pilot-in-command judgment is final.

What is in a pack

Six uint8 planes per 256×256 tile, plus the tile's DEM provenance byte:

plane	what it is
class	surface class (water, wetland, developed, canopy, crop, range, barren…)
conf	confidence in that class, scaled by the source's own confidence band
slope	median slope, 0–50.8°
rough	detrended surface roughness (residual σ)
hazard	fused obstacle/hazard field from towers, roads and transmission corridors
flags	bitfield — charted obstacle, corridor, road buffer, water, wetland, coarse terrain

Facts, never verdicts. Nothing here knows what aircraft is flying. Scoring happens on the device, where the pilot's glide ratio and performance re-tint the same ground — so the same tile means different things to a Cub and to a Malibu, and changing aircraft costs no download.

MBTiles/SQLite containers, z6–z13, planes raw-DEFLATE. index.json at the repo root is the catalog: a flat cells[] array with bounds, size, sha256 and per-source vintages, plus regions[] grouping cells into areas a pilot would recognise.

Overviews never average. Point hazards take the worst child; areal textures take the second-worst of four, because plain max compounds up a pyramid until the map excludes everything (measured: 16% → 73% between z13 and z8 on the first cell). Native resolution is untouched either way.

Coverage

Grows as cells are built. index.json is authoritative — a cell absent from it does not exist. Each entry publishes coarse_terrain_tiles: how much of that cell had only 10 m elevation available, where the ditch- and berm-scale checks cannot run and the score is capped instead.

Sources and attribution

source	licence
USGS 3DEP elevation (1 m where published, else 1/3″)	public domain (US Gov)
USGS Annual NLCD land cover + confidence	public domain (US Gov)
USDA NASS Cropland Data Layer	public domain (US Gov)
USFWS National Wetlands Inventory	public domain (US Gov)
FAA Digital Obstacle File	public domain (US Gov)
US Census TIGER/Line roads	public domain (US Gov)
Meta / WRI canopy height	CC-BY-4.0 — attribution required

Canopy height data © Meta Platforms, Inc. and World Resources Institute, used under CC-BY-4.0. Source imagery is from 2016, so canopy here is roughly a decade old and reads low against present-day growth. Every other input is United States federal public-domain data.

No OpenStreetMap data is present in these tiles

Deliberately, and it is load-bearing rather than incidental. OSM is ODbL, and conflating ODbL geometry into the hazard plane would make this a Derivative Database — a one-way door that could compel opening data these tiles are intended to be combinable with. The hazard plane is built solely from public-domain federal sources, and the build pipeline proves it: the planes are rebuilt with the OSM artifact deleted and compared bit-for-bit.

Verifying a pack

Files are stored as classic Git-LFS, not Xet, because the app fetches them with a plain anonymous GET. If you re-host these, keep it that way:

curl -sL -o n33w107.lzpack \
  https://huggingface.co/datasets/SingularityUS/commsight-lz/resolve/main/cells/n33w107.lzpack
sqlite3 n33w107.lzpack "SELECT name, value FROM metadata WHERE name LIKE 'lz_%';"

The sha256 in index.json is the authority on integrity.