HyperRrs / README.md
TalhaMah56's picture
Dataset card
2a6fb03 verified
|
Raw History Blame Contribute Delete
3.74 kB
metadata
license: cc-by-4.0
pretty_name: HyperRrs aligned multispectral-hyperspectral ocean colour pairs
tags:
  - ocean-color
  - remote-sensing
  - hyperspectral
  - PACE
  - MODIS
  - VIIRS
  - OLCI
size_categories:
  - 100M<n<1B

HyperRrs: aligned multispectral-hyperspectral ocean colour pairs, benchmark, weights and results

Release accompanying the anonymous CVPR 2027 submission "HyperRrs: An Aquatic Foundation Model for Hyperspectral Ocean Color Analysis and Reconstruction".

Each pair is one 1 km cell observed by a multispectral ocean colour sensor (MODIS-Aqua, VIIRS Suomi-NPP, VIIRS NOAA-20 or Sentinel-3 OLCI) and by the hyperspectral PACE OCI, as Level-2 remote-sensing reflectance (Rrs, sr^-1). The PACE target is the OCI Level-2 AOP product, processing version 3.2, on its 172-band grid (346-719 nm).

Contents

folder what size
heritage/{MODISA,VIIRSN,VIIRSJ1}/ aligned pool, one .npz per PACE scene: 2,948 / 3,433 / 3,516 scenes, 128.9 M / 337.2 M / 352.2 M cells (global 1 km equal-area bins, sensor and PACE at most 3 h apart) 302 GiB
olci/ 1,226 OLCI x PACE triplets: 300 m OLCI pixels (olci_300m/), PACE V3.2 targets per role (v32_targets/{val,train400,unused}/), and the earlier 1 km PACE files used to define OLCI test membership (pace_v31_1km_legacy/, not targets) 6.6 GiB
meta/ manifest.csv (one row per scene), matchups.csv (every source granule with its OB.DAAC URL), pace_granules.csv, sanity.json small
counts.json, SHA256SUMS cell / file counts; checksum of every pool file small
benchmark/ the frozen benchmark: test sets in scoring order, training files with scene-group ids and validation masks, OLCI roles, block caches, sensor band operators, leakage checks (README_benchmark.md) 11 GiB
weights/ HyperRrs: pre-trained body and the per-sensor fine-tunes used in the paper's tables (location-and-date and bands-only variants, seeds 0-2), config, normaliser, checksums 3.9 GiB
results/ results.json (every main-table value), stored predictions and per-cell scores for every table row (README.md) 30 GiB

In all, 820,497,483 aligned cells (818.4 M heritage + 2.1 M OLCI).

Pool file schema (heritage/<SENSOR>/<SENSOR>_<PACE time>.npz)

Per cell: input_rrs (n, K) sensor Rrs in its K bands, input_rrs_std, input_n_px, input_prodfail_frac; pace_rrs (n, 172) PACE V3.2 Rrs (NaN = invalid band), pace_rrs_std, pace_n_px, pace_n_bands, pace_row, pace_col; bin_id, lat, lon, dt_s (sensor minus PACE time, s), sensor_granule_idx. Per file: pace_wavelengths_nm, input_wavelengths_nm, input_bands, input_srf_centroid_nm, sensor_granules, pace_granule, sensor_processing_version, flags_used_sensor, flags_used_pace, flags_summary.

The benchmark (benchmark/) is a fixed subsample of this pool with frozen splits; train and score on it to compare with the paper. The pool is released so that other subsamples, regions and periods can be built.

Use

import numpy as np
z = np.load('heritage/MODISA/MODISA_20240306T001405.npz')
X, Y = z['input_rrs'], z['pace_rrs']          # (n, 10), (n, 172)

The HyperRrs inference package and the benchmark runner are distributed with the code (see the paper's release link).

Licence and attribution

Data, benchmark, weights and results: CC BY 4.0. Code: MIT. Derived from NASA PACE OCI, MODIS-Aqua and VIIRS Level-2 ocean colour products (NASA Ocean Biology DAAC; NASA Earth-science data policy: free, full and open). OLCI-derived files: contains modified Copernicus Sentinel data 2024-2025.

Not included: the PACE pre-training store (available from OB.DAAC), Sentinel-2 MSI and EMIT pairs.