MikCil/gargano-sentinel

Gargano Sentinel-2 (clear-sky)

Cloud-filtered, fully-tiled Sentinel-2 acquisitions over the Gargano peninsula (Apulia, Italy). Each acquisition is split into a regular grid of 40 tiles (5 rows × 8 columns) covering roughly 10 × 10 km each at 1000 × 1000 px.

Only complete acquisitions (all 40 tiles available from a single overpass) whose mean tile cloud fraction is ≤ 0.5 are included.

Splits

The dataset is organised in three splits, one per band combination. Every split contains the same (timestamp, row, col) triplets — they differ only in the imagery channel.

Split	Bands	Use-case
rgb	B04, B03, B02 (true-color)	Visual inspection, VLM input
swir_nir_red	B12, B08, B04 (false-color)	Vegetation, fires, geology
rededge	B08, B05, B04	Vegetation health, chlorophyll

Schema

Column	Type	Description
id	string	Unique tile identifier (filename stem)
timestamp	string	ISO-8601 acquisition datetime (UTC)
platform	string	sentinel-2a / sentinel-2b
row	int8	Tile row, 0..4
col	int8	Tile column, 0..7
cloud_fraction	float32	Per-tile cloud cover, 0..1 (Sen2Cor SCL)
valid_pixel_fraction	float32	Fraction of usable pixels
mgrs_codes	string	Source MGRS tile(s)
image	Image	PNG, 1000 × 1000

The id has the form

<YYYYMMDDTHHMMSS>_<s2a|s2b>_r<rr>_c<cc>

e.g. 20230415T100021_s2a_r02_c03. Saving images with this id as the filename makes them directly compatible with the helper scripts below.

Quick start

from datasets import load_dataset

ds = load_dataset("MikCil/gargano-sentinel")
print(ds)            # DatasetDict with rgb / swir_nir_red / rededge splits
print(ds["rgb"][0])  # one tile + metadata

Save all RGB tiles of a given timestamp to disk:

import os
from datasets import load_dataset

ds = load_dataset("MikCil/gargano-sentinel", split="rgb")
day = ds.filter(lambda r: r["timestamp"].startswith("2023-04-15"))

out = "tmp/rgb_2023-04-15"
os.makedirs(out, exist_ok=True)
for ex in day:
    ex["image"].save(f"{out}/{ex['id']}.png")

Helper scripts

Two helper scripts live in scripts/. Both take a folder of tiles named <prefix>_r<rr>_c<cc>.png (the convention used by id above) and group them by prefix automatically — so you can throw multiple timestamps in the same folder if you want.

stitch_full.py — full Gargano image

Stitches all 40 tiles into one full-resolution mosaic (no labels, no borders).

python scripts/stitch_full.py tmp/rgb_2023-04-15
# → tmp/rgb_2023-04-15/stitched_full/<prefix>_full.png   (8000 × 5000 px)

stitch_squares.py — VLM-friendly composite squares

Tiles in their native form are convenient because you can compose them however you like. We additionally provide seven pre-defined squares that keep VLM inputs reasonably sized while preserving full spatial detail:

Square	Range	Size
3 × 3	r00–r02 × c00–c02	3000²
3 × 3	r00–r02 × c03–c05	3000²
2 × 2	r00–r01 × c06–c07	2000²
2 × 2	r02–r03 × c06–c07	2000²
2 × 2	r03–r04 × c00–c01	2000²
2 × 2	r03–r04 × c02–c03	2000²
2 × 2	r03–r04 × c04–c05	2000²

The two south-easternmost tiles (r04c06, r04c07) are pure water and are intentionally not covered.

python scripts/stitch_squares.py tmp/rgb_2023-04-15
# → tmp/rgb_2023-04-15/stitched_squares/<prefix>_3x3_r00-02_c00-02.png
#   tmp/rgb_2023-04-15/stitched_squares/<prefix>_2x2_r03-04_c04-05.png
#   …

License

Sentinel-2 data are made available under the Copernicus Sentinel data licence. This curated derivative is released under CC BY 4.0.

Citation

@dataset{gargano_sentinel2_2026,
  author = {Ciletti, Michele},
  title  = {Gargano Sentinel-2 Dataset},
  year   = {2026},
  url    = {https://huggingface.co/datasets/MikCil/gargano-sentinel}
}