README.md

---
license: cc-by-nc-4.0
task_categories:
- image-to-image
- unconditional-image-generation
- feature-extraction
language:
- en
tags:
- satellite-imagery
- remote-sensing
- hierarchical
- benchmark
- mapbox
- pyramid-generation
size_categories:
- 10K<n<100K
pretty_name: Dense500
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
dataset_info:
  features:
  - name: image
    dtype: image
  - name: location
    dtype: string
  - name: site_id
    dtype: string
  - name: tag
    dtype: string
  - name: continent
    dtype: string
  - name: site_lon
    dtype: float64
  - name: site_lat
    dtype: float64
  - name: zoom
    dtype: int32
  - name: tile_x
    dtype: int32
  - name: tile_y
    dtype: int32
  - name: site_start_zoom
    dtype: int32
  - name: site_leaf_zoom
    dtype: int32
  - name: site_depth
    dtype: int32
  - name: is_root
    dtype: bool
  - name: is_leaf
    dtype: bool
  - name: zoom_window
    dtype: int32
  - name: site_index_in_window
    dtype: int32
  splits:
  - name: train
    num_bytes: 1087489040
    num_examples: 42500
  download_size: 1087489040
  dataset_size: 1087489040
---

# Dense500

**500-site hierarchical satellite imagery benchmark for evaluating pyramid generation across multiple zoom windows.**

500 sites × 85 tiles each (≈42,500 tiles), structured as 100 sites per each of 5 zoom windows (W1–W5). Designed as the evaluation corpus for the `unified` pyramid-generation strategy described in `evaluate_pyramid.md`.

Two distinguishing properties vs sibling
[`global-dense-satellite`](https://huggingface.co/datasets/dcher95/global-dense-satellite):

1. **Fixed depth-4 quadtree per site** — exactly 85 tiles (1 + 4 + 16 + 64) at zooms `l, l+1, l+2, l+3`. No variable-depth sites.
2. **Source-shift filtered** — sites where the imagery source changes mid-hierarchy (e.g., Mapbox satellite → aerial transition around z15–z16 in NA/EU) are **dropped**. Every published site shows visually coherent imagery across all four zooms.

Like `global-dense-satellite`, sites are also disjoint from
[Git-10M](https://huggingface.co/datasets/lcybuaa/Git-10M).

## Zoom windows

| Window | l  | h = l+3 | Root tile covers | Sites |
| --- | --- | --- | --- | --- |
| W1 | 11 | 14 | ~78 km | 100 |
| W2 | 12 | 15 | ~39 km | 100 |
| W3 | 13 | 16 | ~20 km | 100 |
| W4 | 14 | 17 | ~10 km | 100 |
| W5 | 15 | 18 | ~5 km | 100 |

Each site contributes 85 tiles → **42,500 tiles total**.

## What's in here

Each row is one satellite tile. Group by `site_id` to recover a complete depth-4 quadtree; use `zoom_window` to filter by window.

| Column | Type | Description |
| --- | --- | --- |
| `image` | image | 256×256 JPEG bytes |
| `location` | string | `"{zoom}_{x}_{y}"` (Web Mercator) |
| `site_id` | string | Stable id grouping all 85 tiles of one quadtree |
| `tag` | string | `naeu_urban` or `global_urban` |
| `continent` | string | NA, EU, Asia, Africa, SA, Oceania |
| `site_lon`, `site_lat` | float64 | Site center coords |
| `zoom`, `tile_x`, `tile_y` | int32 | Web Mercator tile coords |
| `site_start_zoom`, `site_leaf_zoom` | int32 | Bounds of this site's hierarchy (l, h) |
| `site_depth` | int32 | Always 4 |
| `is_root`, `is_leaf` | bool | Tile position in its hierarchy |
| `zoom_window` | int32 | 1..5 corresponding to W1..W5 |
| `site_index_in_window` | int32 | 0..99 within each window |

## Source-shift filter

A site is dropped if any adjacent-zoom pair within its window has:

```
Bhattacharyya(hist(z), hist(z+1)) > 0.25  AND
Pearson(hist(z), hist(z+1))      < 0.875
```

Histograms are 8×8×8 RGB, aggregated across all tiles at each zoom within the site. Thresholds were calibrated by spot-check on a 50-site W5 probe to sit halfway between a loose default (≥0.30 / <0.85) and a tight alternative (≥0.20 / <0.90).

This filter catches Mapbox's well-known satellite→aerial composite transition in NA/EU cities, which we verified is also present in the upstream Git-10M Google Maps source. The trade-off: about 11–24% of pre-filter sites are dropped, mostly at deeper zoom windows.

## Geographic distribution

| Window | Strategy | Why |
| --- | --- | --- |
| W1 (z11→z14) | Global, all continents | Satellite imagery is reliable globally at these zooms |
| W2 (z12→z15) | Global, all continents | Same, with mild aerial-shift risk near z15 |
| W3 (z13→z16) | NA/EU-biased, other continents represented | z16 outside NA/EU often runs into Mapbox overzoom |
| W4 (z14→z17) | Heavily NA/EU | z17 needs Mapbox's aerial layer (NA/EU only) |
| W5 (z15→z18) | NA/EU only | Same reason — z18 is aerial, satellite would be overzoomed |

## Loading

```python
from datasets import load_dataset

ds = load_dataset("dcher95/dense500", split="train")

# Group tiles by site to recover one full quadtree
from collections import defaultdict
sites = defaultdict(list)
for row in ds:
    sites[row["site_id"]].append(row)

# Or pick all sites in a single zoom window
w3 = ds.filter(lambda r: r["zoom_window"] == 3)
```

For training/eval, use the `GlobalDenseSatelliteDataset` class — its
`(zoom, tile_x, tile_y)` based descendant logic works against this schema
unchanged.

## Limitations and caveats

- **Mapbox tile imagery**: redistribution of Mapbox tiles is restricted by Mapbox's Terms of Service. This dataset is published under CC-BY-NC for non-commercial research use only.
- **NA/EU concentration in deeper windows**: W4 and W5 are heavily NA/EU because that's where Mapbox aerial imagery exists at z17/z18. W1 and W2 are global; W3 has reasonable global coverage but is also NA/EU-biased.
- **Source-shift filter is a heuristic**: it uses color histograms, so a same-source hierarchy with strong seasonal lighting drift across zooms could occasionally trigger a false positive. We checked borderline cases manually during threshold calibration.

## Pipeline

The full generation pipeline is at
[github.com/dcher95/genesis](https://github.com/dcher95/genesis) (paths
`scripts/dataset/dense500/`):

1. `sample_sites_dense500.py` — 5 windows × continent-stratified candidate sampling.
2. `download_tiles.py --full` — concurrent fetch with overzoom validation + site integrity (shared with global_dense).
3. `filter_source_shifts.py` — middle-threshold cross-zoom histogram filter.
4. `topup_window.py` — iterative top-up for underfilled windows (different seed per wave).
5. `finalize_dense500.py` — cap each window at 100, add `zoom_window` + `site_index_in_window` columns.
6. `convert_to_parquet.py`, `publish_to_hf.py` — packaging.

## Citation

```bibtex
@misc{dense500_2026,
  title = {Dense500: A 500-site hierarchical satellite imagery benchmark for pyramid generation evaluation},
  author = {Cher, Dan},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/dcher95/dense500},
}
```