Structured-Noise Deconvolution Benchmark — AMSR-E SST (Full Release)

Anonymous deposit for the NeurIPS 2026 Evaluations & Datasets Track double-blind review. Access is restricted during the review period; reviewers may request access via OpenReview or this deposit page. Will be set to Open upon acceptance.

A representative ~76 MB sample for quick inspection (no access request needed) is available at the companion open-access deposit: DOI 10.5281/zenodo.20044225

CONTENTS (~16.7 GB total)

real_amsre_modis.zip (~5.22 GB) — Co-located AMSR-E / MODIS Level-2 SST patches on a 10 km grid, 32×32 patches. Train: 191,626 patches (2003–2004 and 2006–2011). Test: 25,971 patches (2005, day + night, orbit-based split). SHA-256: 4db78921909f014803041eb2019a7cc6a8101ab98bc28f83fbea05b1f63b7130

simulated_subset.zip (~10.69 GB) — LLC4320-derived simulated AMSR-E / MODIS pairs for the sim-to-obs experiment. SHA-256: 7891d46ea60ffa4de20c89e53d97ba1521188a818b9afcb117d93c6765ccaa05

checkpoints.zip (~803 MB) — 138 trained model checkpoints (.pth) covering every reported run, organized by experiment name. SHA-256: a9fe208988d154ece2cb373b03fec87b76b8e941c3cf0c72ba7a373f57d85c07

zenodo_manifest.sha256 (467 B) — Per-file SHA-256, size, and record counts for the three archives above.

VERIFYING THE DOWNLOAD

Reviewers can verify all files at once with: shasum -a 256 -c zenodo_manifest.sha256

CODE AND REPRODUCTION

Training code, evaluation scripts, configuration files, and reproduction notebooks are linked from the OpenReview submission (https://anonymous.4open.science/r/structured-noise-deconvolution-benchmark-84DB/)

METADATA

Protocol version: canonical-2026-04-20 (frozen). License: CC-BY-4.0 for derivatives. Users must respect the upstream licenses of source assets: AMSR-E (Remote Sensing Systems), MODIS-Aqua (NASA OB.DAAC), and LLC4320