TalhaMah56/pace-s2-emit-sr-tiles

PACE × Sentinel-2 × EMIT super-resolution training tiles

18,126 tiles cut from the pace-s2-emit-triplets scenes, with the frozen train/val/test split those experiments were scored on.

Tiles are cut at sub_window = 24, stride = 24 (no overlap) in PACE pixels: each tile is 24 × 24 PACE pixels and the co-located 2400 × 2400 Sentinel-2 pixels — a ×100 spatial ratio.

split	tiles
train	12,573
val	2,627
test	2,926
total	18,126

The tiles are re-cuttable from the parent scenes with a short script. The split is not — manifest.json records which tile went where, and every published number from the super-resolution track was scored against exactly that partition.

Contents per file (~32 MB)

key	shape	dtype
pace_rrs	(172, 24, 24)	float16
pace_validity	(24, 24)	uint8
s2_10m_refl / s2_10m_validity	(4, 2400, 2400) / (2400, 2400)	int16 / uint8
s2_20m_refl / s2_20m_validity	(6, 1200, 1200) / (1200, 1200)	int16 / uint8
s2_60m_refl / s2_60m_validity	(2, 400, 400) / (400, 400)	int16 / uint8
emit_rfl / emit_validity	(45, 400, 400) / (400, 400)	float16 / uint8
provenance	triplet_id, granule_name, win_r0, win_c0, sub_window
location / time	lat_centroid, lon_centroid, pace_year, pace_doy

triplet_id + win_r0 + win_c0 joins each tile back to its parent scene, so the two datasets are usable together. Wavelength vectors and band-name arrays are not stored per tile — take them from the parent scene.

Known limitations — read before use

• Inherits every limitation of the parent dataset: the PACE leg is product version V3_1 (withdrawn from OB.DAAC), Sentinel-2 is uncorrected L1C top-of-atmosphere, EMIT is L2A surface reflectance, and the Sentinel-2 and EMIT legs are not radiometrically comparable as stored.
• The split is by tile, not by scene. Tiles from one parent scene appear in more than one split. Group by triplet_id before quoting any generalisation number — leakage through exactly this route has produced wrong answers in this project before.

What it was built for

Training data for a PACE → Sentinel-2 spatial super-resolution investigation that closed negatively: cross-sensor spatial super-resolution over water did not work, encoder-inclusively. The split is published so the negative result can be reproduced rather than taken on trust.

Attribution

Derived from NASA PACE OCI L2 AOP (V3_1) and NASA EMIT L2A RFL (v001), courtesy of the NASA Ocean Biology DAAC and the LP DAAC, and from Copernicus Sentinel-2 L1C data (2022–2025).