Research map / Detection and segmentation / Open-set and few-shot detection
MFogHub: Bridging Multi-Regional and Multi-Satellite Data for Global Marine Fog Detection and Forecasting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces MFogHub, a multi-regional, multi-satellite dataset (about 68,000 samples from 15 coastal regions and 6 geostationary satellites) with a cube-stream structure, and benchmarks 16 detection and forecasting baselines to assess regional and satellite generalization.
Problem
Marine fog detection and forecasting research relies on open datasets that are limited to a single region or satellite, which restricts evaluation of generalization across diverse conditions.
Approach
A multi-regional, multi-satellite benchmark organized as region-satellite "cube-streams" (timestamp x spectral band x lat x lon), with pixel-level fog masks for detection and sequence targets for forecasting, used to measure cross-region and cross-satellite generalization.
Key results
- DlinkViT achieves the best cross-satellite detection performance on both FY4A and H8/9 in the Yellow and Bohai Sea region (CSI 56.93 on FY4A, 60.25 on H8/9).
- Multi-region training consistently outperforms single-region training for detection, particularly for the B.C. region.
- Natural-color band combinations outperform true-color combinations for detection on H8/9 and FY4A, and H8/9 yields better results than FY4A.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train on combined B.C., C.C., and G.A. regional data instead of one region | marine fog detection models · B.C., C.C., and G.A. sub-datasets from GOES | CSI | Helped |
| train on closely related combined regional data instead of a single region | TAU · MeteoSat forecasting sub-datasets | MSE | Helped |
| use H8/9 satellite data instead of FY4A satellite data | marine fog detection and forecasting baselines · Yellow and Bohai Sea | CSI | Mixed |
| use natural-color bands instead of true-color bands (ablation) | marine fog detection models · Yellow and Bohai Sea, H8/9 and FY4A | CSI: 51.27 -> 58.10; 50.06 -> 54.84 | Helped |
| use all available spectral bands instead of a three-band combination (ablation) | marine fog detection models · Yellow and Bohai Sea | CSI | Helped |
| increase the negative-sample proportion from 1:1 to 2:1 positive-to-negative (ablation) | marine fog detection models · B.C. and G.A. sub-datasets from GOES | CSI | Mixed |
| remove negative samples, using a 1:0 positive-to-negative ratio (ablation) | marine fog detection models · B.C. and G.A. sub-datasets from GOES | CSI | Mixed |
Surprising findings
- Simpler architectures and ViT-based models show the most promise for multi-region training, and ViT-based models perform well on B.C. and G.A. but weakly on the simpler C.C. sub-dataset.
- Forecasting models show stronger consistency across regions than detection models, and PhyDNet ranks highest in PSNR for M.W. but worst for Na.
Limitations
- Only about 11,600 of 68,000 samples are pixel-annotated, relying on meteorological reports for labels.
- Multi-region experiments are restricted to a single satellite because channel counts differ across satellites.
Experimental setup
Datasets: MFogHub ICOADS
Benchmarks: MFogHub
Metrics: CSI recall precision mAcc mIoU MSE MAE SSIM PSNR
Baselines: Deeplabv3+ UNet UNet++ ViT DlinkViT Unetformer BANet ABCNet ConvLSTM PredRNN MIM PhyDNet
Builds on: OpenSTL
Closest papers by meaning
- CoastalBench: A Decade-Long High-Resolution Dataset to Emulate Complex Coastal Processes (ICML 2025 · similarity 0.85)
- FuXi-Ocean: A Global Ocean Forecasting System with Sub-Daily Resolution (NeurIPS 2025 · similarity 0.84)
- IceDiff: High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior (CVPR 2025 · similarity 0.83)
- OxyGenerator: Reconstructing Global Ocean Deoxygenation Over a Century with Deep Learning (ICML 2024 · similarity 0.83)
- Terra: A Multimodal Spatio-Temporal Dataset Spanning the Earth (NeurIPS 2024 · similarity 0.83)
- Early Detection of Extreme Storm Tide Events Using Multimodal Data Processing (AAAI 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.