TalhaMah56/PACE_v3_1

PACE OCI L2 AOP V3_1 pre-training corpus

Tiled remote-sensing reflectance (Rrs) patches from NASA PACE OCI Level-2 Apparent Optical Properties, packed as .tar shards, built for masked-autoencoder pre-training on hyperspectral ocean colour.

This is product version V3_1, which NASA has withdrawn. OB.DAAC returns 404 on every V3_1 route and CMR now serves only version 3.2. This corpus cannot be rebuilt from the archive, which is the reason it is published here.

Why the version matters

The V3_1 → V3_2 reprocessing is not cosmetic. Measured on 310k common pixels across 8 dual-version granules:

quantity	change
median Rrs change per band	3.3 %
spectral shape change	1.03° median SAM (mean 1.74°, p90 3.92°)
worst bands	−10.4 % at 346 nm, −9.9 % at 706 nm, +4.0 % at 661 nm

That is the same magnitude as typical clear-water model margins. A model trained on this corpus and evaluated on V3_2 data carries a silent train/test radiometry mismatch. Every current and future PACE file a user obtains from NASA is V3_2 or later. Treat this corpus as a historical artefact and a reproducibility record, not as the basis for a model you intend to ship against live PACE data.

⚠️ This corpus is split across two repositories

The corpus did not fit in one Hugging Face account's storage, so it is split in two. The shards never overlap and the layout is identical: copy train/ and val/ from both repos into one directory and you have the whole corpus.

repository	holds
TalhaMah56/PACE_v3_1 (this repo)	train/shard_14503.tar … train/shard_24202.tar: 9,700 shards, byte-verified. No val/.
TalhaMah56/PACE_v3_1_part2	train/shard_00000–00359 and train/shard_24203–24497 (655 shards). No val/.

The full corpus is 24,498 train shards (shard_00000 … shard_24497) and 1,332 val shards (shard_00000 … shard_01331). The two repos together hold 10,355 train shards and no val. Not uploaded: train/shard_00360–14502 (14,143 shards) and all of val/. Those still exist offline; open a discussion if you need them.

Other datasets from the same project

repository	what it is
TalhaMah56/PACE_v3_1_part2	part 2 of this corpus
TalhaMah56/pace-s2-emit-triplets	PACE × Sentinel-2 × EMIT scenes on a shared grid
TalhaMah56/pace-s2-emit-sr-tiles	super-resolution training tiles cut from those scenes

Layout

Each split is a directory of .tar shards named shard_NNNNN.tar. The split is held out by granule hash (5 %), so no acquisition appears in both — which also means the validation split cannot be reconstructed by subsetting the train shards here.

Each shard is a tar of patches. Per patch: Rrs as int16 with a scale factor, a one-bit per-pixel validity mask, and scalar metadata (lat/lon, year/day-of-year, solar zenith, flag fractions, per-granule quality terms). 172 bands.

Known limitations — these are measured, and they matter

Publishing this corpus without its defects would be misleading. All of the following were measured on the corpus itself:

• The quality score measures missing pixels, not contamination. patch_score is dominated by hard_frac (LAND + PRODFAIL + cloud-fill), 75 % of the median score, and correlates with 1 − valid_frac at r = 0.966. Because turbid water sits next to land and under cloud, turbid tiles are down-weighted ~20 % relative to the rest — the sampling works against the hardest water.
• Composition is overwhelmingly clear water. Tiles with ≥10 % TURBIDW pixels: 1.6 %. COASTZ ≥10 %: 3.4 %. Coccolithophore: 0.8 %. By latitude: 35 % within |lat|<20°, only 2.7 % above 60°. By year: 2024 39 % / 2025 52 % / 2026 9 %.
• Redundant along the wrong axis. Tiles are cut at stride 64 on 128-px patches, so every pixel appears in up to four tiles: ~7.8 M patches for roughly 1.1 M tile-equivalents of unique valid area.
• Invalid pixels are stored as zero. ~12 % of pixels inside loss-bearing tokens are fabricated zeros and 58 % of such tokens contain at least one. Use the validity mask; never treat zero as a measurement.
• Side information was discarded at build time. The per-pixel l2_flags word, Rrs_unc, aot_865, angstrom, avw and nflh are not in the shards. This matters more than it sounds: among valid ocean pixels, 36.5 % are flagged STRAYLIGHT, 34.8 % ABSAER, 31.6 % MODGLINT, and 71.2 % lie within 8 px of an invalid pixel — and nothing in the stored data lets a loss function know.
• Red-band noise. On clear water the product's own propagated uncertainty is ~48 % relative at 670 nm and ~52 % at 719 nm (vs ~5 % at 443 nm). The red bands are close to half noise.
• Negative reflectance from the NIR atmospheric correction. ~4 % of valid pixels are negative at 670 nm on clear granules; over turbid and inland water the over-subtraction is worse.
• No processing_version field is stored in the corpus. The version is known only externally — from this card.

Provenance

Built from the PACE_OCI_L2_AOP V3_1 archive (2024-03-05 → 2026-03-01), 97,341-granule catalogue, ~99 % of which was retrieved. Tiling, quality scoring and sharding were done by a project-internal pipeline.

Attribution

NASA PACE OCI Level-2 Apparent Optical Properties (V3_1), courtesy of the NASA Ocean Biology Distributed Active Archive Center (OB.DAAC). PACE data are NASA public-domain products.