README.md

metadata

license: cc-by-4.0
dataset_info:
  features:
    - name: solar_images
      dtype: image
    - name: xrs
      dtype: float32
  configs:
    - config_name: default
      data_files:
        - split: train
          path: '**/*'
language:
  - en
tags:
  - solar-physics
  - space-weather
  - flare-prediction
  - astronomy
  - time-series
  - computer-vision
  - deep-learning
  - iccv2025
size_categories:
  - 10K<n<100K
task_categories:
  - time-series-forecasting
  - image-classification
pretty_name: FlareBench

FlareBench: A Comprehensive Benchmark for Solar Flare Prediction 🌞

FlareBench is a novel benchmark dataset for solar flare prediction that covers the entire 11-year solar activity cycle. This dataset was introduced in our ICCV 2025 paper "Deep Space Weather Model: Long-Range Solar Flare Prediction from Multi-Wavelength Images".

🚀 What's New

• ICCV 2025 Accepted! Our paper has been accepted to ICCV 2025
• Complete Solar Cycle Coverage: First benchmark covering a full 11-year solar activity cycle
• Multi-Modal Data: Combines solar images (AIA/HMI) and X-ray sensor data (XRS)
• Time-Series Cross-Validation: Designed for robust evaluation across diverse solar activity states

📊 Dataset Overview

FlareBench addresses the limitations of conventional solar flare prediction datasets that often:

• Cover only limited periods of solar activity
• Contain small, low-resolution sunspot patches
• Exhibit biases towards specific solar activity periods

Why FlareBench?

Most conventional datasets for solar flare prediction do not cover diverse solar activity states. Consequently, models trained on such datasets can exhibit biases towards specific periods of solar activity. Furthermore, many datasets contain only small, low-resolution sunspot patches.

FlareBench presents unique challenges:

• Requires modeling long-term, diverse solar states spanning the entire 11-year solar cycle
• Demands computationally efficient architectures for multi-wavelength images capturing multi-layered physical phenomena
• Handles highly imbalanced class distributions that vary significantly across different years

🗂️ Dataset Structure

FlareBench/
├── solar_images/
│   ├── aia/          # Atmospheric Imaging Assembly (9 wavelengths)
│   │   ├── 2011/     # 📊 Data available (Apr-Dec 2011)
│   │   ├── 2012/     # 📁 Directory structure only
│   │   ├── ...
│   │   └── 2024/
│   └── hmi/          # Helioseismic and Magnetic Imager
│       ├── 2011/     # 📊 Data available (Jan-Dec 2011)  
│       ├── 2012/     # 📁 Directory structure only
│       └── ...
└── xrs/              # X-Ray Sensor data from GOES satellites
    ├── 2011/         # 📊 Data available (full year)
    ├── 2012/         # 📁 Directory structure only
    └── ...

📈 Dataset Statistics

• Total Samples: 95,837 (after quality filtering)
• Time Period: June 2011 - November 2022 (full solar cycle)
• Temporal Resolution: 1-hour cadence
• Spatial Resolution: Full-disk solar observations
• Class Distribution:
  • X-class: 1,750 samples (1.8%)
  • M-class: 13,263 samples (13.8%)
  • C-class: 34,978 samples (36.5%)
  • No-flare: 47,775 samples (49.9%)

Data Composition

Solar Images:

• AIA (Atmospheric Imaging Assembly): Multi-wavelength extreme ultraviolet observations capturing the multi-layered coronal atmosphere
• HMI (Helioseismic and Magnetic Imager): Line-of-sight and vector magnetic field observations from the photosphere

X-Ray Sensor Data:

• XRS: Solar X-ray flux measurements from GOES satellites
• Formats: Both CSV and NetCDF files
• Content: Raw and processed solar X-ray flux data

🎯 Prediction Task

FlareBench focuses on predicting the maximum class of solar flare within the next 24 hours, following standard approaches in solar flare prediction research. The prediction classes are:

• X-class: Major flares (≥10⁻⁴ W/m²)
• M-class: Moderate flares (10⁻⁵ to 10⁻⁴ W/m²)
• C-class: Minor flares (10⁻⁶ to 10⁻⁵ W/m²)
• No-flare: Below C-class threshold

📧 Request Full Dataset

Need data from other years (2012-2022)? Please contact us at: [email protected]

We will provide access to the complete dataset for research purposes. Please include:

• Your research affiliation
• Brief description of your research project
• Intended use of the dataset
• Preferred data transfer method

🔬 Evaluation Protocol

FlareBench uses time-series cross-validation to ensure unbiased evaluation:

• 3-fold cross-validation covering different phases of the solar cycle
• Chronological splits maintaining temporal order
• Diverse solar conditions in each fold's test set
• Standard metrics: TSS, BSS, GMGS for comprehensive evaluation

📚 Citation

If you use FlareBench in your research, please cite our ICCV 2025 paper:

@inproceedings{nagashima2025deepswm,
  title={Deep Space Weather Model: Long-Range Solar Flare Prediction from Multi-Wavelength Images},
  author={Shunya Nagashima and Komei Sugiura},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  year={2025}
}

🔗 Related Resources

• Paper: arXiv:2508.07847
• Project Page: https://keio-smilab25.github.io/DeepSWM
• Code: https://github.com/keio-smilab25/DeepSWM

📄 License

This dataset is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

🤝 Acknowledgments

This dataset is built using observations from:

• Solar Dynamics Observatory (SDO): NASA's flagship solar observation mission
• GOES Satellites: NOAA's Geostationary Operational Environmental Satellites

📞 Contact

For questions, issues, or collaboration opportunities:

• Primary Contact: Shunya Nagashima - [email protected]
• Institution: Keio University
• Project Page: https://keio-smilab25.github.io/DeepSWM

---

Built with ❤️ for the solar physics and machine learning communities