Jingjing Wang
State Key Laboratory of Solar Activity and Space Weather, National Space Science Center, Chinese Academy of Sciences, Beijing 100190, China

Operational space-weather forecasting starts with a deceptively simple question: what is happening on the Sun right now? Answering it requires identifying active regions (ARs), coronal holes (CHs), and filaments in a flood of solar images, linking them to official region numbers, and connecting today’s magnetic field to tomorrow’s flares. Ref. [1] released the Solar Situation Awareness and Flare Forecasting Dataset, which converts 14 years of SDO/HMI observations into exactly such ready-to-use products, with a historical extension reaching back to 1996.

Figure 1. Standardized Solar Observation Image Dataset. Full-disk examples from SDO/HMI, SDO/AIA, and ground-based Hɑ, all centered and rescaled to a common field of view extending to 1.1 solar radii.

The core dataset covers May 2010 to May 2024 and combines HMI line-of-sight magnetograms and continuum images with SDO/AIA EUV images at 131, 193, 211, and 304 Å and ground-based Hɑ observations. A historical extension adds SOHO/MDI magnetograms and SOHO/EIT 195 and 304 Å images from January 1996, extending the standardized magnetic record to nearly three decades. All images are rescaled to a common solar-disk coordinate system reaching 1.1 solar radii and organized on a unified daily index, providing spatially consistent, co-registered observations across instruments (see Figure 1).

Figure 2. Schematic diagram of manually annotated NOAA ARs based on SolarMonitor AR data and SDO/HMI full-disk magnetogram (DATE=20230507). (a) Daily solar AR image released by SolarMonitor website; (b) SDO/HMI full-disk magnetogram; (c) Manually annotated AR image, where differently colored rectangular regions and numbers denote distinct NOAA ARs.

A central contribution of the dataset is its long-term collection of expert-reviewed annotations. For every officially numbered NOAA AR, experts used daily NOAA/SWPC Solar Region Summaries and SolarMonitor information together with HMI magnetograms, continuum and AIA images to identify the magnetic structure, which was then refined into spatially coherent masks through adaptive segmentation and subsequently inspected and corrected by experts before release. The released annotations are therefore curated products of human identification, image processing, and expert checking, rather than raw outputs of an automated recognition model. Figure 2 illustrates this workflow for NOAA ARs. Beyond ARs, the same expert-driven framework also covers coronal holes and filaments: CHs were delineated mainly from AIA 193 Å images with reference to NOAA synoptic maps, and 3919 Hɑ images were manually annotated for filaments.

The dataset also resolves a long-standing correspondence problem between HMI research products and operational region numbering: SHARP patches and NOAA AR numbers do not always correspond one-to-one. The team manually mapped 1945 non-one-to-one SHARP regions involving 2394 NOAA ARs, providing an explicit link between the two systems.

Each daily AR is further characterized by operational descriptors: total unsigned magnetic flux, AR area, the R-value measuring flux concentrated near polarity inversion lines (PILs), mean and PIL-weighted field gradients, Hale and McIntosh classes, and a flare index for the preceding 24 hours. Notably, although the released images are clipped to ±500 G for display and machine learning, all parameters are computed from the original, unclipped HMI FITS data. Three independent labels then record whether each AR produces at least one C-, M-, or X-class flare within the subsequent 24 hours.

These descriptors prove physically meaningful in two tests. Total unsigned flux agrees well with SHARP values for ARs near disk center. Among 2,005 evaluated AR samples, the fraction producing at least one C-class-or-above flare within the subsequent 24 hours (232 samples, 11.6% overall) rises steadily with every magnetic parameter, consistent with the monotonic trend expected for flare-relevant parameters.

The dataset has already been applied end to end: it underpins the Solar Activity AI Forecaster^[2], a framework that integrates solar situation perception, physical-parameter analysis, and strong-flare forecasting. Its perception module was trained on these expert annotations and then scaled through a human-in-the-loop cycle in which experts inspected and corrected model-labeled samples before retraining, combining machine scalability with expert quality control.

The complete dataset, including standardized observations, expert-reviewed AR/CH/filament masks, the SHARP–NOAA mapping, magnetic parameters, and 24-hour flare labels, is publicly available at https://doi.org/10.57760/sciencedb.space.03541. It provides a reusable foundation for solar-structure recognition, multimodal machine learning, and operational flare forecasting. For more details, please refer to Ref. [1].

References

[1] Li, M., Liu, S., Wang, J., et al., 2026, Sci Data, doi: 10.1038/s41597-026-08005-5.

[2] Wang, J., Liang, P., Wang, T., et al., 2025, arXiv e-prints, arXiv:2508.06892, doi: 10.48550/arXiv.2508.06892.