Solar Active Region Detector (SARD)

• SARD is a family of automated systems that detect and characterize solar active regions from various solar observations, improving space-weather monitoring.
• It employs advanced image segmentation and machine learning techniques, using inputs like magnetograms, EUV, and continuum images to capture magnetic flux concentrations even without visible sunspots.
• The approach, including YOLOv8-based detectors and clustering methods, achieves high accuracy and offers actionable insights for predicting flares, CMEs, and other eruptive solar events.

Solar Active Region Detector (SARD) denotes automated systems for identifying solar active regions, that is, regions where strong magnetic fields concentrate at the photosphere, often corresponding to sunspots and serving as the source of flares, coronal mass ejections, and other eruptive events. In the literature, the term refers both to a specific YOLOv8-based detector for full-disk SDO/HMI line-of-sight magnetograms and to a broader class of SARD-like pipelines that detect, segment, track, catalog, and characterize active regions from magnetograms, continuum images, EUV imagery, helioseismic products, or derived time series for space-weather monitoring and forecasting (Pan et al., 29 Jul 2025, Pérez-Suárez et al., 2011).

1. Definition and observational scope

A solar active region is physically a concentration of magnetic flux emerging through the solar surface, but a consistent operational definition is difficult because the appearance of an active region depends strongly on wavelength and atmospheric layer. The review literature states that a consistent definition is “somewhat difficult, and occasionally controversial,” and explicitly treats active-region detection as one component of a broader automated solar feature detection framework rather than as an isolated task (Pérez-Suárez et al., 2011).

This observational dependence is central to SARD design. In white-light or continuum images, active regions are represented by sunspot groups in the photosphere; in magnetograms, they are magnetic-flux concentrations; in H-alpha or other chromospheric images, they can appear as bright patches; and in EUV images, they can appear as high-contrast loop-like structures. The same region is therefore not identical across observables, and the literature repeatedly emphasizes that there is no single detection method valid across all image types (Pérez-Suárez et al., 2011).

The space-weather motivation follows directly from this physical definition. Active regions are the source of major eruptive phenomena, especially flares and coronal mass ejections, so region-based forecasting requires determining the magnetic properties of an active region and its surroundings. A recurring misconception is that active-region detection is equivalent to sunspot counting. Magnetogram-based systems instead target magnetic structure directly, which allows detection of magnetically active regions that may not have a white-light signature and may therefore be missed in traditional cataloging schemes tied to visible sunspots (Pérez-Suárez et al., 2011).

2. Historical lineage and rule-based foundations

The principal historical precursor to modern SARD systems is SMART, the SolarMonitor Active Region Tracking system. SMART is an automated system that detects, segments, tracks, catalogs, and characterizes active regions from magnetograms, and the review literature identifies it as the closest early match to a SARD-like active-region detector. Its four-step logic is explicit: segment magnetograms into individual feature masks, characterize each extracted region, classify regions, and catalog and track them through time (Pérez-Suárez et al., 2011).

The SMART detection core is a preprocessing and segmentation pipeline based on two consecutive magnetograms. The magnetogram is smoothed with a 2D Gaussian, a static threshold is used to remove the background, pixels above threshold are set to 1 to create a binary mask, the masks are radially dilated, the dilated masks from consecutive times are subtracted to identify transient features, and the resulting mask is dilated again before region characterization. The review gives thresholding at 70 G as the example shown in Figure 1 and notes that the use of two consecutive magnetograms removes transient features and supports temporal analysis from first emergence through reappearance over multiple solar rotations (Pérez-Suárez et al., 2011).

The detailed SMART description adds the operational reasons for this architecture. SMART relies on consecutive image differencing to remove both quiet-Sun and transient magnetic features, uses region-growing techniques to group flux concentrations into classifiable features, determines magnetic properties such as region size, total flux, flux imbalance, flux emergence rate, Schrijver’s $R$-value, $R^*$, and Falconer’s measurement of non-potentiality, and uses a persistence algorithm to associate developed active regions with emerging flux regions in previous measurements and to track regions beyond the limb through multiple solar rotations (Higgins et al., 2010).

Comparative work on SMART, ASAP, STARA, and SPoCA established several methodological baselines that remained influential for later SARD variants. Magnetogram-based detection was presented as the most stable foundation for long-term active-region tracking; white-light detectors were more tightly tied to visible sunspot structure; and EUV segmenters were useful for coronal context but more prone to merging neighboring regions into a single coronal structure. The same comparison concluded that summed area is more stable and less ambiguous than the number of detected features, and that magnetic flux and total sunspot area are the best indicators of active-region emergence (Verbeeck et al., 2011).

3. Data modalities, benchmark datasets, and learning-ready representations

SARD workflows span several input modalities, but magnetograms are repeatedly identified as the most informative input for space-weather use because active regions are magnetic-flux concentrations. This emphasis on magnetic measurements shaped later benchmark datasets, particularly those intended not merely for full-disk detection but for the downstream stages of classification, tracking, and flare prediction (Pérez-Suárez et al., 2011).

A major example is the Solar Active Region Magnetogram Image Dataset, which provides SDO/HMI line-of-sight magnetograms of solar active regions together with flare labels. The dataset is assembled from three sources: NOAA Solar Region Summaries define when a NOAA active region is visible on the solar disk and provide latitude and longitude; SDO/HMI images are downloaded through JSOC as $600 \times 600$ pixel active-region-centered cutouts tracked at the Carrington rotation rate; and NOAA Event Reports provide flare labels, including peak flare time and GOES class. Each image receives either a flare-class label such as C1.0, M4.7, or X10.1, or the label 0 for no flare, with the label set by the largest flare in the prediction window. In the preconfigured dataset, the default is a 24-hour prediction window with a minimum flare threshold of C1.0. The dataset is split at the active-region level into 1,256 regions for training, 157 for validation, and 157 for testing, and it also includes 29 magnetic complexity features grouped into 7 gradient features, 13 neutral-line features, 5 wavelet features, and 4 flux features (Boucheron et al., 2023).

The same dataset formalizes standard flare-prediction evaluation. Using the 29-feature table with an SVM employing a linear kernel and balanced class weights, the reported test-set results for 24-hour flare occurrence are $\mathrm{TPR}=0.7484$, $\mathrm{TNR}=0.7791$, $\mathrm{HSS}=0.4485$, and $\mathrm{TSS}=0.5275$ for full resolution, with similar performance for reduced-resolution data. A transfer-learning experiment using VGG16, with the final classification layer replaced by a two-class output and only the last fully connected layer fine-tuned, yields performance comparable to the SVM results. These results situate SARD in a broader supervised and unsupervised machine-learning setting rather than restricting it to full-disk segmentation alone (Boucheron et al., 2023).

SolARED extends the SARD problem from established active regions to active-region emergence prediction. It is derived from full-disk SDO/HMI observations of Doppler velocity $V_d$, line-of-sight magnetic field or unsigned magnetic flux $\Phi_B$ or $\Phi_m$, and continuum intensity $R^*$0, and spans March 1, 2010 to June 1, 2023. The dataset covers 50 large active regions that emerged within $R^*$1 of the central meridian, persisted for more than 4 days, and reached a total area of 200 millionths of a solar hemisphere. For each region, a $R^*$2 patch corresponding to $R^*$3 pixels is tracked for 10 days, remapped into heliographic coordinates using Postel’s projection, divided into a $R^*$4 grid, and normalized to reduce center-to-limb effects. Across all 50 active regions, 42/50 regions, or 84%, show magnetic flux increase before continuum-intensity decrease, with an average lead time of 16.25 hours and a standard deviation of 18.53 hours. The dataset is therefore explicitly structured for early-warning studies of active-region emergence rather than only for established-region recognition (Kasapis et al., 19 Jan 2026).

4. Algorithmic realizations

Later work diversified SARD into object detection, density-based clustering, hierarchical clustering with tracking, helioseismic segmentation, and above-limb semantic segmentation.

System	Core method	Primary domain
SARD	YOLOv8 with DCN, ECA, and an extra $R^*$5 detection head	Full-disk SDO/HMI magnetograms
DSARD	Thresholding plus two-stage DBSCAN clustering and merging	Full-disk SDO/HMI magnetograms
HARDAT	HDBSCAN with differential-rotation tracking and PIL extraction by SVC	SOHO/MDI and SDO/HMI magnetograms
Deep-learning farside detector	U-net on helioseismic phase-shift maps	Solar far side
304 Å limb detector	U-Net semantic segmentation	Above-limb SDO/AIA 304 Å images

The named SARD system is the YOLOv8-based detector trained on 6,975 SDO/HMI full-disk line-of-sight magnetograms from 7 April 2010 to 31 December 2019, sampled at a 12-hour cadence, with 26,531 manually refined active-region labels. Its pre-labeling pipeline uses Otsu threshold segmentation, morphological opening with an $R^*$6 square structuring element, region growing with an intensity threshold of 0.05, and morphological closing; the magnetograms are saturated to $R^*$7, and pixels beyond the solar limb are set to zero. The detector is built on YOLOv8s, with a backbone, neck based on FPN and PAN, and detection head, and adds deformable convolution in the C2f block, Efficient Channel Attention, and an additional $R^*$8 small-object head, so that the final detector operates on four scales rather than three (Pan et al., 29 Jul 2025).

DSARD, by contrast, is an unsupervised implementation centered on DBSCAN. It applies threshold-based image segmentation at 150 Gauss, a first global DBSCAN with $R^*$9 pixels and $600 \times 600$0 pixels, a second reformative DBSCAN for oversized clusters exceeding $600 \times 600$1 with $600 \times 600$2 pixels, and an integration stage using polarity categories defined by a ratio of 10, a merge distance of 50 Mm, and size filtering at 70 $600 \times 600$3 for neutral clusters and 300 $600 \times 600$4 for positive or negative clusters. Its final output is a set of active regions with minimum bounding rectangles, positions, areas, magnetic fluxes, and bipolar tilt angles (Chen et al., 25 Feb 2025).

HARDAT is an explicit extension of DSARD. It replaces the fixed-density DBSCAN core with HDBSCAN, applies threshold segmentation at 100 G for MDI and 150 G for HMI, uses HDBSCAN cluster selection to handle multi-density magnetic structures, follows that with a light DBSCAN cleanup, and adds a multi-object tracking stage based on solar differential rotation and Hamming-distance similarity. It also introduces a polarity inversion line extractor, PIL-svc, based on support vector classification with an RBF kernel using $600 \times 600$5 and $600 \times 600$6. In this formulation, SARD becomes a detect-track-analyze framework rather than only a detector (Shi et al., 22 May 2026).

Two specialized branches further extend the concept. On the far side of the Sun, deep learning has been used to map 11 consecutive helioseismic phase-shift maps into a probability map for active-region presence via a U-net, replacing fixed seismic-strength thresholding with an integrated probability criterion $600 \times 600$7 (Felipe et al., 2019). Above the solar limb, a U-Net-based three-class semantic segmentation model detects active regions, prominences, and background in SDO/AIA 304 Å images after conversion to polar coordinates, rescaling, and three-channel tone mapping, thereby shifting SARD from magnetogram segmentation to off-limb EUV morphology (Zhang et al., 2024).

5. Outputs, validation, and catalog generation

The outputs of SARD systems vary by method, but the review literature identifies a common set: feature masks or segmentation masks, contours of the active region, physical properties such as size, total magnetic flux, flux imbalance, growth or decay rate, magnetic morphology, and downstream catalog entries with temporal tracks. In SMART, the segmentation output is a mask used to plot contours, and the later characterization step computes the physical properties of each extracted region (Pérez-Suárez et al., 2011).

The YOLOv8-based SARD outputs bounding boxes around active regions and evaluates predictions with an intersection-over-union threshold of 50%. Its reported overall performance is $600 \times 600$8, precision $600 \times 600$9, recall $\mathrm{TPR}=0.7484$0, and $\mathrm{TPR}=0.7484$1. On the hardest high-density test set T1, the paper reports $\mathrm{TPR}=0.7484$2, precision $\mathrm{TPR}=0.7484$3, recall $\mathrm{TPR}=0.7484$4, and $\mathrm{TPR}=0.7484$5. The ablation study reports AP $\mathrm{TPR}=0.7484$6 for baseline YOLOv8s, AP $\mathrm{TPR}=0.7484$7 after adding the extra head, AP $\mathrm{TPR}=0.7484$8 after adding ECA, and AP $\mathrm{TPR}=0.7484$9 for the full SARD model with DCN. The same work uses 1,331 active regions within $\mathrm{TNR}=0.7791$0 of the central meridian to study area and unsigned magnetic flux and finds that both are best described by log-normal distributions (Pan et al., 29 Jul 2025).

DSARD is validated against NOAA by processing 4,984 full-disk magnetograms and producing 25,178 active-region detections over the full disk; after restriction to $\mathrm{TNR}=0.7791$1 around the central meridian, 2,863 regions are used for validation. The reported average true positive rate is 91.8%, and the average false positive rate is 7.2%. The paper emphasizes that the detector can identify smaller or more diffuse active regions missed by NOAA and can separate nearby regions that region-growing methods tend to merge (Chen et al., 25 Feb 2025).

For far-side active-region detection, validation is necessarily indirect. Comparison of HMI helioseismic far-side detections with STEREO/EUVI 304 Å observations found that 100% of the 22 seismic far-side regions correspond to an EUV plage at essentially the same location, and 95% correspond to an EUV plage that either became a NOAA-designated magnetic region upon reaching the east limb or had been a NOAA-designated region before crossing to the far side. The same study reports no false far-side active-region detections among those 22 seismic regions when the improved five-day HMI processing is used (Liewer et al., 2017).

Above the limb, the U-Net detector reports for active regions precision $\mathrm{TNR}=0.7791$2, recall $\mathrm{TNR}=0.7791$3, $\mathrm{TNR}=0.7791$4, and IoU $\mathrm{TNR}=0.7791$5 on a 22-image test set. Applied to one image per day from 2010 May 13 to 2020 December 31, it identifies 10,018 active regions and produces solar-cycle statistics, including the result that about 99.1% of detected active regions lie below $\mathrm{TNR}=0.7791$6 latitude (Zhang et al., 2024).

6. Space-weather applications, operational forecasting, and limitations

The space-weather role of SARD is explicit throughout the literature. Active regions are the source of flares and coronal mass ejections, and region-based flare and CME forecasting requires determining the magnetic properties of an active region and its surroundings. The review literature further notes that future versions of SMART may incorporate flare event probability using magnetic flux near polarity separation lines, following Schrijver (2007), proxies for non-potentiality following Falconer et al. (2008), and more advanced descriptors such as fractal techniques, multifractal techniques, and 2D wavelet transform modulus maxima methods (Pérez-Suárez et al., 2011).

An operational extension of this logic appears in the physics-based eruptive-region monitor built from observed line-of-sight magnetograms, 3D data-driven magnetofrictional simulations, and forward projection in time. After a ramp-up phase of about 35 magnetograms, or about 2 days, the method projects forward for 10 magnetograms, corresponding to about 10–16 hours, and computes an operational eruption-risk score $\mathrm{TNR}=0.7791$7. It assigns three warning levels: red for $\mathrm{TNR}=0.7791$8, amber for $\mathrm{TNR}=0.7791$9, and green for $\mathrm{HSS}=0.4485$0. In the reported proof of concept, 4 out of 5 eruptive regions are successfully classified for the relevant pre-eruption interval, and the three non-eruptive regions generally settle to green as the forecast progresses (Pagano et al., 2019).

The limitations of SARD are correspondingly varied. The underlying definition of an active region remains wavelength-dependent and sometimes controversial; NOAA assigns numbers only to regions with a white-light signature, so magnetically active regions without sunspots may be missed in traditional cataloging; thresholding and segmentation depend strongly on instrumentation and observing conditions; higher-resolution observations may make features better defined but also harder to detect because previously contiguous regions may fragment; and ground-based observations suffer from atmospheric instability (Pérez-Suárez et al., 2011). Dataset-specific caveats add further constraints. The solar active-region magnetogram benchmark does not correct for projection effects and therefore restricts the default preconfigured dataset to within $\mathrm{HSS}=0.4485$1 latitude and longitude while removing images containing NaNs (Boucheron et al., 2023). SolARED is intentionally restricted to 50 large active regions within $\mathrm{HSS}=0.4485$2 of the central meridian, and its thresholds are expressed in non-dimensional normalized units (Kasapis et al., 19 Jan 2026). HARDAT, while more adaptive than DSARD, is also computationally heavier because of the HDBSCAN core and the added tracking and PIL-extraction modules (Shi et al., 22 May 2026).

Taken together, the literature presents SARD both as a named YOLOv8 detector and as a family of space-weather-oriented systems that combine preprocessing, segmentation, characterization, tracking, cataloging, and, in some cases, forecasting. The common thread is not a single algorithmic template but a magnetically informed strategy for turning heterogeneous solar observations into stable active-region inventories and operationally useful measures of emergence, evolution, and eruptive potential (Pan et al., 29 Jul 2025, Pérez-Suárez et al., 2011).