2026-10-14T06:03:42Z https://www.scidb.cn/oai

10.57760/sciencedb.012p4 2026-09-18T11:31:00Z

2026-09-18 A 20-city benchmark for single-channel nighttime CLDMSK emulation from VIIRS Day/Night Band imagery doi:10.57760/sciencedb.012p4 en Nighttime cloud masking normally relies on thermal infrared channels, which many visible-only small-satellite missions do not carry. This dataset provides a reproducible cross-city benchmark for the question of whether a single VIIRS Day/Night Band (DNB) channel can reproduce the operational VIIRS CLDMSK cloud-mask decision. Collocating DNB Level-1 radiance with CLDMSK Level-2 within 1 km over 2023–2024 yields patches from 20 cities across 712 distinct overpass nights; array-level quality screening retains 14,149 paired 200×200-pixel patches (53 removed for scan-edge fill bands, CLDMSK "no result" pixels, or fill lunar metadata). Under a strict leave-one-city-out protocol with five seeds (100 trained models), a radiance-only ResNet34 U-Net reaches 0.8079 mean pixel-level Emulation Agreement (EA), with city means spanning 16.8 percentage points (Sydney 0.723 to Cairo 0.892) against a seed-to-seed standard deviation of 0.003. Non-learned baselines fall far behind (majority class 0.563, single-pixel radiance threshold 0.552, pixel-statistics random forest 0.639), and three paired ablations — illumination/geometry conditioning via FiLM (−0.73 pp), per-city z-score normalisation (+0.21 pp), and explicit local-uniformity channels (−0.13 pp) — all fail to improve cross-city transfer. The deposit contains the collocation index and exclusion list, the fold definitions and seeds, all per-fold result files, the preprocessing and evaluation code, and the sources of every figure and table, so that every number in the associated paper can be reproduced. Contents: ScienceDB_deposit_core_20260914.zip (code, benchmark index, per-fold results, figures, tables, documentation) and dataset_20city_clean.zip (14,149 .npz patches, layout <city>/<City>_YYYYMMDD_HHMMSSS.npz; field dictionary in DATA_DICTIONARY.md). nighttime cloud detection; VIIRS Day/Night Band; CLDMSK; emulation agreement; cross-city transfer; leave-one-city-out; deep learning; benchmark dataset mingyu chen shensen hu weihua ai shuo ma PUBLIC https://creativecommons.org/licenses/by/4.0/ dataset Science Data Bank