Datasets:

whereyaheading
/

copernicus-glo30-h3-terrain

How this dataset is checked for correctness, and how to stress-test it for your own use. Everything here is reproducible from the public dataset + source; the checks are organized cheapest-first. Each item lists what it verifies and a pass criterion.

TL;DR for a quick trust check: run the structural invariants (Tier 1) over a sample, then the landmark spot-checks (Tier 2). If those pass, the dataset is behaving.

Tier 1 — Structural invariants (every cell, cheap)

Asserted over the whole dataset (or a random sample per resolution). All must hold for 100% of rows.

Check	Pass criterion
Column order & types	`h3_cell uint64`, four `elevation_* float32`, `pixel_count uint32`, `geoid_undulation float32`
Ordering	`min ≤ mean ≤ max` (within fp tolerance)
Non-negativity	`elevation_stddev ≥ 0`, `pixel_count ≥ 1`
Finiteness	no NaN / Inf in any column
Physical range	every `elevation_*` ∈ [−432 m (Dead Sea shore), 8849 m (Everest)]
Geoid range	`geoid_undulation` ∈ [−110 m, +90 m] (EGM2008 global bounds)
Cell validity	every `h3_cell` is a valid H3 index at the file's resolution
Uniqueness	`h3_cell` is unique within the dataset for a given resolution

-- one-liner range/ordering check on res 8 (DuckDB)
SELECT count(*) AS violations FROM 'res8/base=*/*.parquet'
WHERE NOT (elevation_min <= elevation_mean AND elevation_mean <= elevation_max
           AND elevation_stddev >= 0 AND pixel_count >= 1
           AND elevation_max BETWEEN -432 AND 8849);   -- expect 0

Tier 2 — Landmark spot-checks (ground truth)

Look up known points and compare. Flat/open sites → check elevation_mean; sharp peaks → check elevation_max (a 30 m DSM under-samples summits, so expect the Copernicus value, not the survey height).

Site	(lat, lon)	Column	Expected	Tol
LAX airport	33.9416, −118.4085	mean	~38 m	±25
Death Valley (Badwater)	36.2468, −116.8143	mean	~−86 m	±15
Four Corners	36.999, −109.045	mean	~1477 m	±20
Aspen (KASE)	39.2232, −106.8687	mean	~2380 m	±40
Everest summit	27.9881, 86.9250	max	~8738 m (≈111 m below survey)	±60

Geoid spot-checks (compare geoid_undulation to the NGS / online EGM2008 calculator):

Site	Expected N
KTEB (Teterboro)	−32.8 m
LAX	−36.0 m
KDEN (Denver)	−18.2 m
London	+46.5 m
Tokyo	+36.7 m

End-to-end AGL check (the headline use): for a known on-ground ADS-B fix, height_above_ground = alt_geom − elevation_mean − geoid_undulation should be ≈ 0. Worked: KTEB −29 − 1 − (−32) ≈ +2 m. ✓

Tier 3 — Cross-source comparison

Sample N random cells and compare elevation_mean against an independent source:

Raw Copernicus DEM at the same lat/lon — should match within floating-point precision (this is the round-trip check; a mismatch means an aggregation bug).
USGS 3DEP (US cells only) — an independent high-accuracy DEM; expect agreement within DEM error (a few m), larger over canopy/buildings (DSM vs DTM).
Kontur 400 m H3 elevation — close, modulo their aggregation choices.

Pass: median absolute difference vs raw Copernicus ≈ 0; vs 3DEP within ~few m (bias explainable by DSM-vs-DTM).

Tier 4 — Internal consistency

Boundary-merge correctness. A hex that straddled a processing seam must equal a single-pass aggregation over the union of its pixels — exactly on pixel_count/min/max, fp-exact on mean. (Validated: e.g. a straddling res-9 hex with partials of 41 px @ 430 m and 89 px @ 456 m → merged 130 px @ 447.94 m — the pixel-weighted mean, not the naïve average.)
Conservation. Σ pixel_count over a resolution == total valid source pixels processed (nothing lost or double-counted).
Cross-resolution consistency. Rolling res-(k+1) cells up to their res-k parents should approximately match the directly-computed res-k file (gap expected from H3's imperfect nesting; a large divergence flags a bug).
Partition integrity. Every cell in res{8,9}/base=B/ actually has base cell B ((h3_cell >> 45) & 127 = B). Expect 0 misfiled cells. (No H3 library needed — the shard key is pure bit math.)
No-orphan / completeness. Every land tile in the source produced cells; emitted land extent reconciles against an independent landmask. (Safety-relevant: a missing cell is interpreted as ocean / sea level, so dropped land would silently read as 0 m.)

# partition-integrity spot check (no h3 extension required)
import duckdb
bad = duckdb.sql("""
  SELECT count(*) FROM read_parquet('res9/base=2/*.parquet')
  WHERE (h3_cell >> 45) & 127 <> 2
""").fetchone()[0]   # expect 0

Tier 5 — Distributional checks

Hypsographic curve. The global distribution of elevation_mean (area-weighted by cell) should match the shape of Earth's known hypsographic curve — most land low-lying, a long tail into the mountains — with sane fractions below 200 m / above 1000 m. This catches systematic bias a spot-check can't.

Observed (this build): the area-weighted mean of elevation_mean over res-5 cells is ~612 m, with 44 % of cells below 200 m and ~20 % above 1000 m. That is lower than the ~800–840 m figure often quoted for "mean land elevation." We believe the gap is a property of what this dataset measures, not an aggregation defect, for three reasons: (1) it is a DSM area-weighted mean — res-5 H3 cells are near-equal-area, so a plain cell average already is area-weighted, and there is no double-counting (the conservation check confirms the same pixel total at every resolution); (2) the dataset is land-only and flat-lines large water bodies at ~0 m, which drags the mean down relative to figures computed differently; and (3) the canonical "840 m" number is itself source- and method-dependent and especially sensitive to how the high Antarctic/Greenland ice sheets are weighted. So we read 612 m as the honest area-weighted mean of this surface model, and treat the hypsographic check as a shape/sanity test (low-skewed with a mountain tail) rather than a hard match to a literature constant. If a future consumer needs a true bare-earth (DTM) land mean, this number should not be used as that figure.
Geoid field shape. geoid_undulation should be smooth and large-wavelength: negative across CONUS, positive over Europe/Japan, no high-frequency noise.
Coverage by latitude. Cell counts per latitude band should track land area (the dataset is land-only).

Stress & performance testing

Targets for "will this hold up under my workload."

Test	What it measures	Healthy result
Point lookup, cold	latency to read one res-9 cell via shard pruning	reads ~~1 base shard (~~130 MB), not the whole dataset; sub-second after the file is cached
Shard pruning	a regional query touches only the relevant `base=` shards	files scanned ≈ #base cells in the query box, not all 111
Bulk join	join N million of your points against res-8 on `h3_cell`	linear in N; memory bounded (project to `h3_cell` first)
Full-resolution scan	aggregate over res-9 (1.7 B rows)	completes streaming on a laptop (DuckDB), bounded memory
Repartition / re-export	rebuild a different partition scheme	the res-9 GROUP-BY needs ~100 GB scratch — do it in-RAM (≥128 GB) or with ample disk

Edge cases to exercise:

Ocean lookup → no row returned; confirm your code treats "missing" as sea level (0 m orthometric), not as an error.
Sub-sea-level land (Dead Sea ~−430 m, Death Valley ~−86 m) → negative elevation_mean present and correct.
High-elevation fields (KASE ~2380 m, KDEN ~1655 m) → the AGL formula needs geoid_undulation; alt_baro/ground flags alone are wrong here.
Poles & dateline → cells near 84 °N and −90 °S, and at ±180° longitude, resolve and look up correctly.
Shard-seam cells → a cell whose neighbors fall in a different base= shard is still found via its own base cell.

Reproducibility

The dataset is a deterministic derivative of public inputs. A rebuild from the same Copernicus GLO-30 vintage + the pinned H3 / PROJ versions (recorded in each Parquet footer) reproduces the cell values bit-for-bit on count/min/max and within fp on mean/stddev. The build + these checks live in this repository.

Known limitations

See the README caveats: it's a DSM (surface, not bare ground), land-only/sparse (missing = sea level), under-reports sharp peaks, flat-lines water, and carries the usual polar DEM artifacts.