YBSF-55: Reference Annotations for Daytime Sea Fog Detection over the Yellow and Bohai Seas

Abstract

Dataset Overview The Yellow and Bohai Sea Fog dataset (YBSF-55) contains 417 reference annotations from 55 selected daytime observation dates during 2016–2020. It covers sea fog over the Yellow and Bohai Seas, including scenes with partial cloud obscuration, within 30–42°N and 117–129°E. The annotations correspond to Himawari-8 Advanced Himawari Imager (AHI) observations provided by the Japan Meteorological Agency (JMA) and accessed through NICT Science Cloud, Japan. The observations used in the study were calibrated, geolocated, regionally extracted, and resampled onto a common 600 × 600 grid with 16 spectral channels and approximately 2 km nominal sampling. Original satellite observations are obtained separately from the providers. Data-access information and the preprocessing overview are provided in README.md. Observations span 03:00–05:20 UTC (11:00–13:20 China Standard Time) at a nominal 10-min cadence. Each annotation corresponds to the latest observation, at target time t, in a three-frame sequence acquired at t−20 min, t−10 min, and t. Twenty-four dates have all 13 eligible targets from 03:20 to 05:20 UTC; the other dates have partial or more sparsely sampled annotations. Exact annotation and input-scan timestamps are provided in the metadata. Missing frames were not interpolated when constructing temporal samples. Two experienced meteorological specialists manually delineated sea fog and cloud using multiband composites of AHI B03–B05 and temporal context. Sea fog and cloud were drawn as separate layers in Adobe Photoshop, with their overlap defining cloud-obscured sea fog. A fixed land mask identifies land, and remaining ocean areas are classified as clear ocean. The five reference categories are land, clear ocean, sea fog, cloud, and cloud-obscured sea fog. The released annotations are 1350 × 1350 RGBA PNG reference visualizations generated from categorical labels with display processing and geographical overlays. Their export dimensions differ from the 600 × 600 regional arrays used in the study. Image filenames identify the annotated target scan in UTC. Dataset contents ground_truth1-6.zip : The 417 reference annotations corresponding to the 55 selected observation dates. observation_dates.csv: Observation dates, annotation counts, and first and last annotated target times. ground_truth_alignment.csv: Annotation filenames, target timestamps, and corresponding three-frame AHI scan timestamps. README.md: Data-source access, study region, preprocessing overview, annotation categories, and metadata definitions. The references represent expert interpretation of the imagery and temporal context, including contextual interpretation of sea fog beneath overlying clouds. Associated research code is available separately at https://github.com/TangYuzhu/SITS-SFNet.