TalhaMah56/pace-s2-emit-triplets

PACE × Sentinel-2 × EMIT co-registered water scenes

2,859 scenes, each a compressed .npz holding one PACE OCI hyperspectral granule, one to four Sentinel-2 L1C tiles and one EMIT L2A scene resampled onto a single common UTM grid, with a per-sensor validity mask. Median footprint 117 × 114 km; ~1.2 GB per file; 2.7 TiB total.

Same-day hyperspectral-to-multispectral water scenes at 1 km / 60 m / 10 m on one grid — the co-registration is the asset. The source granules are public; the alignment cost a multi-day pipeline run (6.5 h per 429 scenes) that no longer exists in runnable form.

Contents per file

key	shape	dtype	note
pace_rrs	(172, ~117, ~120)	float16	PACE OCI L2 AOP remote-sensing reflectance, 1 km
pace_validity	(~117, ~120)	uint8
pace_wavelengths_nm	(172,)	float32
emit_rfl	(45, 1950, 2000)	float16	EMIT L2A surface reflectance, 60 m, 45-band subset
emit_validity / emit_wavelengths_nm / emit_band_idx_native			native band indices kept for traceability
s2_10m_refl	(4, 11700, 12000)	int16	B2 B3 B4 B8
s2_20m_refl	(6, 5850, 6000)	int16	B5 B6 B7 B8A B11 B12
s2_60m_refl	(2, 1950, 2000)	int16	B1 B9
s2_*_validity, s2_*_band_names			per resolution
grid	utm_epsg, grid_left, grid_top, grid_w_m, grid_h_m		the common grid
provenance	triplet_id, granule_name, pace_datetime, emit_datetime, s2_datetime, pace_emit_dt_hr, pace_s2_dt_hr, includes_emit		acquisition offsets in hours
location / time	lat_centroid, lon_centroid, pace_year, pace_doy

Sidecars: _ledger.json (3,484 catalogue groups → 2,561 ok / 923 skipped, with valid_frac_pace/emit/s2 and the PACE partner name per scene), _summary.json, _validation_report.json, _failures.csv, _corrupt_files.csv, dropable_rows.csv.

Known limitations — read before use

• The PACE leg is product version V3_1 in every scene. That version has since been withdrawn from OB.DAAC and cannot be obtained from NASA any more. The V3_1 → V3_2 reprocessing changes Rrs by a median 3.3 % per band and ~1.03° SAM (worst at band edges: −10.4 % at 346 nm, −9.9 % at 706 nm, +4.0 % at 661 nm). Anyone using the PACE leg quantitatively needs to know this — and it is also why this copy is worth publishing.
• PACE coverage is often thin. _validation_report.json flags 872 of 2,875 scenes with issues, most of them pace_rrs is >95 % NaN. The report ships with the data; gate on it rather than assuming every scene is usable.
• Sentinel-2 is L1C top-of-atmosphere int16, not atmospherically corrected, while EMIT is L2A surface reflectance. The two legs are not radiometrically comparable as stored.
• _ledger.json and _summary.json contain Windows paths (D:\, F:\) from the machine that built them. Harmless, but they are not valid paths for a user.

What it was built for

This is the training corpus of a PACE → Sentinel-2 spatial super-resolution investigation that closed negatively: cross-sensor spatial super-resolution over water did not work, including with the encoder in the loop, and the chained variants through EMIT and Sentinel-3 closed too. Publishing the corpus is what makes that negative result checkable.

A companion dataset of 24×24-PACE-pixel training tiles cut from these scenes, with the frozen train/val/test split used for those experiments, is published separately.

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).