dcher95/global-dense-satellite

global_dense

Globally distributed, dense hierarchical satellite imagery for evaluating hierarchical satellite generation models.

20,608 satellite tiles across 256 sites worldwide, structured as complete parent-child quadtrees from zoom 10 down to zoom 18. Designed as an out-of-distribution evaluation set for models trained on Git-10M — every tile in this dataset is guaranteed disjoint from Git-10M's location set.

What's in here

Each row is one satellite tile. A "site" is a complete dense quadtree (every tile at every zoom between site_start_zoom and site_leaf_zoom). Different sites span different zoom ranges depending on geography:

• NA/EU urban (41 sites): can reach z18 because Mapbox has aerial imagery in North America and Europe. Hierarchies are typically z14→z18 (341 tiles).
• Global urban (157 sites): cities everywhere; capped at z17 where Mapbox satellite is reliable. Hierarchies z14→z17 (85 tiles) or shallower.
• Rural (58 sites): low-population areas worldwide. Shallow z10→z12 or z11→z13 hierarchies (21 tiles).

Statistic	Value
Sites	256
Tiles	20,608
Zoom range	z10 — z18
Distinct depth-from-root values	3, 4, 5 levels
Continents	NA, EU, Asia, Africa, SA, Oceania
Source	Mapbox Satellite (mapbox.satellite, .jpg90, 256×256)

Site center distribution

Site centers were sampled with LandScan 2024 population as a prior, continent-stratified, with minimum 80 km spacing for urban tiers and 120 km for rural. Per-site zoom range is drawn from a distribution conditioned on (in NA/EU?, local population density).

Why not Git-10M overlap

Strict any-tile exclusion: a candidate site was dropped if any tile in its full quadtree (every zoom × every x/y) matched a zoom_x_y location in Git-10M's location_to_indices. This makes the dataset a true held-out evaluation set for Git-10M-trained models.

Source-shift annotation

Mapbox's mapbox.satellite is a composite. In NA/EU at z15+ they typically switch from satellite imagery to high-resolution aerial photography sourced from national mapping agencies. The change produces a visible color/sensor discontinuity between adjacent zoom levels within a site (this is also true of Google Maps Satellite, which is what Git-10M sources from — so models trained on Git-10M have already seen this artifact).

Three columns annotate this per row:

• site_has_source_shift (bool): whether the analyzer detected a cross-zoom source/color discontinuity in this site.
• site_shift_zoom (int32): the zoom level where the shift starts (-1 if no shift).
• is_post_shift (bool): whether this tile's zoom is ≥ the shift zoom — i.e., this tile is from the post-shift (aerial) source.

Detection method: per-zoom 8×8×8 RGB histograms aggregated across all tiles at that zoom within a site, compared pairwise via Bhattacharyya distance and Pearson correlation. Flagged when Bhattacharyya > 0.30 AND correlation < 0.85.

Distribution of flagged sites: 0% of rural, 11% of global urban, 29% of NA/EU urban — consistent with the Mapbox compositing pattern.

Schema

Column	Type	Description
image	image	256×256 JPEG bytes
location	string	Tile location key, format "{zoom}_{x}_{y}" (Web Mercator)
site_id	string	Stable id grouping all tiles of one quadtree
tag	string	naeu_urban, global_urban, or rural
continent	string	NA, EU, Asia, Africa, SA, Oceania
site_lon, site_lat	float64	Site center coords (degrees)
zoom, tile_x, tile_y	int32	Web Mercator tile coords
site_start_zoom	int32	Shallowest zoom in this site's hierarchy
site_leaf_zoom	int32	Deepest zoom in this site's hierarchy
site_depth	int32	leaf - start + 1
is_root	bool	Tile is the single root of its site
is_leaf	bool	Tile is at the deepest zoom of its site
site_has_source_shift	bool	See "Source-shift annotation"
site_shift_zoom	int32	-1 if no shift, else first post-shift zoom
is_post_shift	bool	True iff zoom >= site_shift_zoom (and a shift exists)

Validation

Every tile passed an at-download-time overzoom check (is_blank_or_blurry: reject if grayscale Laplacian variance < 10 OR stddev < 5). Any site with even one failed tile was dropped entirely so the surviving hierarchies are all complete.

Loading

from datasets import load_dataset

ds = load_dataset("dcher95/global-dense-satellite", split="train")
# Or stream for partial reads:
ds = load_dataset("dcher95/global-dense-satellite", split="train", streaming=True)

# Group tiles by site:
from collections import defaultdict
by_site = defaultdict(list)
for row in ds:
    by_site[row["site_id"]].append(row)

Limitations and caveats

• Mapbox tile imagery: redistribution of Mapbox tiles outside personal/ caching contexts is restricted by Mapbox's Terms of Service. This dataset is published under CC-BY-NC for non-commercial research use; if you need commercial use, contact Mapbox.
• Cross-zoom source shift: ~12% of sites (mostly NA/EU cities) show a satellite→aerial source change between z15 and z16. This is in-distribution for Git-10M-trained models but worth being aware of for any model that expects perfect cross-zoom color consistency.
• English continent labels: continent assignment is via bounding boxes, not a true polygon test — some border regions may be labeled approximately.
• Equator-skewed coverage: LandScan-weighted sampling concentrates population centers, which are largely outside high latitudes. Polar / near-polar regions are absent by design.

Pipeline

The full generation pipeline is at github.com/dcher95/genesis (paths scripts/dataset/global_dense/):

1. sample_sites.py — LandScan-weighted candidate sampling with continent stratification and per-tier zoom-range distributions.
2. add_sites.py — supplement an existing manifest for top-ups.
3. download_tiles.py --full — concurrent fetch from Mapbox with site-level integrity (drop a site if any tile fails validation). Supports resume from an existing dataset.
4. analyze_source_shifts.py — per-site cross-zoom histogram analysis → the three flag columns.
5. convert_to_parquet.py, publish_to_hf.py — packaging.

License

CC-BY-NC-4.0. Mapbox imagery is included under this license for non-commercial research evaluation. © Mapbox / © OpenStreetMap contributors / © Maxar / others per Mapbox's source attribution policy.

Citation

@misc{global_dense_2026,
  title = {global_dense: A globally distributed dense hierarchical satellite imagery benchmark},
  author = {Cher, Dan},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/dcher95/global-dense-satellite},
}