{"id":"e3b18229-5d76-454d-b89a-8c99270e181f","arxiv_id":"2507.21866","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"SARD, a YOLOv8 detector with added attention and deformable convolutions, finds solar active regions in magnetograms with about 94% AP, and the areas and fluxes of those regions are found to follow log-normal distributions.","lead":"The paper trains a YOLOv8-based object detector called SARD to find solar active regions in SDO/HMI magnetograms, reporting about 94% average precision on a labeled dataset. A smart generalist might read it as an example of applying modern deep learning object detection to solar physics, with implications for automated space weather monitoring.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported image and label counts are internally inconsistent, so the 94% AP and no-overlap claim cannot be verified as stated.","rationale":"Reading the paper in good faith, the model architecture is a reasonable YOLOv8 variant and the ablation study is plausible. However, the strongest claim is quantitative, and the paper's own numbers do not add up. The sum of image counts in Table 1 (6,863) contradicts the stated 6,975 in the abstract and §2.2, and the AR count changes from 26,531 in §2.2 to 19,098 in §6. This is not a stylistic issue; it means the dataset used for training and evaluation is not specified precisely enough to reproduce the claimed AP. The 'without any overlap' guarantee depends on split definitions that are also unclear for T2. The reader identified the ground-truth labeling as the weakest assumption; I partially agree, but the internal inconsistency is a more direct obstacle: if the numbers do not match, the evaluation is not auditable. The proposed check—re-running the released code and data—would settle this. If the counts match one of the stated totals and the recomputed AP matches Table 4, the concern is resolved. Otherwise, the central claim is not supported. Since the reader already assigned CONDITIONAL, my assessment does not change that verdict.","tokens_in":13120,"tokens_out":4323,"duration_ms":46410,"concrete_test":"Download the released dataset and code (Data Availability URL and GitHub repo), enumerate all magnetogram files, and assign each to train/test splits using the provided manifest or timestamps. Verify (1) total image count equals 6,975 or explain the 112-image shortfall; (2) labeled AR count is 26,531 or 19,098; (3) no test image is within 14 days of any training image; (4) using the released model, recompute AP on T1, T2, and T3 and compare with Table 4. If any discrepancy persists, the headline AP cannot be taken at face value.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is an AP as high as 94% on test sets with no overlap, but the dataset description is internally inconsistent. The abstract and §2.2 state 6,975 magnetograms and 26,531 labeled ARs; Table 1 lists 5,063 training + 311 (T1) + 790 (T2) + 699 (T3) = 6,863 images, 112 fewer. Section 6 then says 19,098 labeled ARs, about 7,400 fewer. Test T2's period ('Entire 2010; 2011.1.1–2011.6.16') at 12-hour cadence would contain roughly 1,792 images, not 790, so the selection criteria are unclear. If the actual test-set composition or ground-truth labels differ from those used to produce Tables 3 and 4, the reported AP of 0.943 may not be reproducible, and the no-overlap guarantee cannot be audited without a manifest. The reader's concern about NOAA-derived ground truth is valid, but these numeric inconsistencies are a more immediate threat: whatever labels are used, the described dataset does not match the stated numbers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents SARD, a YOLOv8-based object detection system for solar active regions (ARs) in SDO/HMI line-of-sight magnetograms. The authors construct a labeled dataset by applying thresholding, morphological operations, and region growing to 6,975 magnetograms from 2010 to 2019 and manually adjusting the automated labels to match NOAA/SolarMonitor AR annotations. They introduce three modifications to the YOLOv8 architecture (an additional small-object detection head, an ECA attention module, and deformable convolutions), train the model with a temporally split training/test protocol, and report precision around 0.89, recall around 0.88, and average precision around 0.94 on three test sets. They further fit log-normal, exponential, and power-law functions to the area and magnetic flux distributions of the detected ARs and conclude that log-normal distributions provide the best fit. The paper provides links to the dataset and code and includes an ablation study of the architectural changes.","tokens_in":13377,"tokens_out":6513,"duration_ms":73035,"significance":"If the reported performance is reproducible, SARD would be a practically useful tool for automated AR detection, and the public release of code and data is a positive contribution. The temporal split with a 14-day gap is a sensible attempt to prevent data leakage, and the ablation study gives partial credit for each architectural modification. However, the current text contains unresolved inconsistencies in dataset counts and metric values that prevent verification of the central 'no overlap' and 'AP as high as 94%' claims, and the statistical analysis of the log-normal result needs more careful fitting and model-comparison methodology. These issues are fixable but currently load-bearing.","major_comments":[{"comment":"The dataset description is internally inconsistent. The abstract and §2.2 state 6,975 magnetograms and 26,531 labeled ARs, but the image counts in Table 1 sum to 5,063 + 311 + 790 + 699 = 6,863, and §6 states 19,098 labeled ARs. In addition, Test 2 is described as 'Entire 2010; 2011.1.1–2011.6.16' at a 12-hour cadence, which would contain roughly 1,000 images rather than the listed 790. Please provide a complete data manifest with per-set image identifiers and label counts, and reconcile all numbers. Without this, the no-overlap guarantee and the reported AP of 0.943 cannot be independently audited.","section":"§2.2, Table 1, and §6"},{"comment":"The performance numbers are inconsistent across the text and tables. The text reports overall precision, recall, F1, and AP of 0.895, 0.884, 0.889, and 0.946, while Table 3 and the 'Overall' row of Table 4 list 0.89, 0.881, 0.886, and 0.943; the T1 values also differ between the text (0.896/0.901/0.898/0.95) and Table 4 (0.893/0.898/0.896/0.947). Please use one consistent set of metrics, state the exact confidence threshold and IoU criterion used for each reported number, and clarify whether AP is computed at IoU = 0.5 only.","section":"§4.2, Table 3, and Table 4"},{"comment":"The statistical analysis needs revision. The log-normal density in Eq. (10) is written with exp(−ln(x/μ)^2/(2σ^2)); with this parameterization μ is a scale parameter, not the expected value as the text states, so the fitted μ values are not directly comparable to the usual log-normal mean. The power-law fit fixes xmin to the median of the data, and the reported KS p-values are computed on the same data used to estimate the parameters; this is not a valid model comparison across distributions fitted over different ranges. Please provide the fitting algorithm, parameter uncertainties, and a fair model comparison (e.g., bootstrap or information criteria) before claiming that log-normal is the best fit. The unsupported assertion in §5 that excluding newly emerging ARs 'barely affected' the statistics should also be quantified.","section":"§5, Eq. (10)"},{"comment":"The ground-truth labels are produced by an automated pipeline and then manually adjusted to match NOAA/SolarMonitor AR annotations, but the paper does not quantify the frequency or extent of manual adjustment, nor does it compare the final labels with an independent AR catalog or estimate inter-annotator variability. Because both the AP values and the flux/area distributions are computed against these labels, the results are conditional on the NOAA AR definition. Please add a quantitative description of the labeling workflow and, ideally, a comparison of the final labels to the raw NOAA positions and sizes.","section":"§2.1 and §2.2"}],"minor_comments":[{"comment":"The caption spells 'OSTU' where the text and standard usage write 'Otsu'; please correct the spelling.","section":"Figure 1 caption"},{"comment":"There are typographical errors: 'vales' should be 'values' in §2.1, and 'phtospheric' should be 'photospheric' in §5.","section":"§2.1 and §5"},{"comment":"The dates '12 October 2010' and '10 October 23' are garbled; AR 11856 belongs to 2013, and Figure 1 shows 12 October 2013, so the dates and figure references should be corrected.","section":"§4.2"},{"comment":"The note 'see Kingma (2014) for AdamW optimizer' is incorrect; AdamW is introduced by Loshchilov and Hutter (2019), while Kingma and Ba (2015) introduced Adam. Please update the citation.","section":"Table 2 note"},{"comment":"When reporting the KS-test results for the full dataset, the paper gives p-values for the flux distribution but does not give the corresponding p-values for the area distribution; please report them consistently for both distributions and all fitted models.","section":"§5"}],"recommendation":"major_revision","confidential_remarks":"The central detection claim is plausible but cannot be verified from the current text because of the unresolved dataset and metric inconsistencies. The statistical section also needs a more rigorous fitting protocol. These issues appear to be reporting and methodology gaps rather than fatal flaws, and the availability of code and data makes correction feasible within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful part of this paper is concrete: a trained YOLOv8 variant for detecting solar active regions in HMI line-of-sight magnetograms, with code and data released. The ablations show each modification (extra head, ECA, DCN) gives a small, plausible improvement, and the final model beats both the plain YOLOv8 baselines and the older threshold/morphology pipeline. If you work on automated AR detection, this is a reasonable reference point.\n\nThe flaws are mostly in the numbers. The abstract says 26,531 labeled ARs; the conclusion says 19,098. The image counts from Table 1 sum to 6,863, not 6,975 as stated. Test set T2's period at 12-hour cadence should contain roughly 1,800 images, not 790. The text reports AP 0.946 and F1 0.889, while Table 4 says 0.943 and 0.886. Any one of these could be a typo, but together they mean the headline '94% with no overlap' cannot be audited from the paper as written. A dataset manifest would settle it.\n\nThe statistical section is weaker. Fixing the power-law xmin to the median effectively guarantees a bad power-law fit, so the KS-test comparison is not informative. And the paper never explains how the ±3.5° central-meridian selection avoids duplicate counting of the same AR across multiple 12-hour frames. The log-normal conclusion is plausible and consistent with Bogdan et al. and Ograpishvili, but it is a confirmatory result, not a new one, and the current fit comparison doesn't prove it definitively.\n\nI also share the reader's worry about the NOAA/SolarMonitor-derived ground truth. The authors do manual adjustment, but they don't quantify how much the automated labeling is changed or how often. That matters because both the AP and the distributions inherit whatever biases are in those labels.\n\nBottom line: this paper deserves a serious referee, but only after the authors clean up the numerical inconsistencies and reframe the statistics as a comparison the data can actually support. I'd send it to review with a request for a dataset manifest and a reanalysis of the power-law fit. If those come back solid, the model itself is a genuinely useful contribution.","headline":"Useful, reproducible YOLOv8-based AR detector, but the paper's numbers don't add up and the statistical comparison is stacked; fixable with a careful revision.","tokens_in":13939,"tokens_out":1839,"would_cite":true,"duration_ms":22695,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A YOLOv8-based detector finds solar active regions in full-disk magnetograms with 94% average precision and under one second per image.","keywords":["solar active regions","object detection","YOLOv8","SDO/HMI magnetograms","deep learning","log-normal distribution","solar cycle 24","deformable convolution"],"falsifier":"Build an independent test set of full-disk magnetograms in which active regions are manually outlined by several solar physicists according to a fixed magnetic-flux definition, deliberately excluding NOAA group boundaries and including weak and decayed regions; compute SARD's average precision at IoU>0.5 on that set. If it falls well below 0.94, the headline generalization fails; if the area and flux histograms from the independently segmented data are not log-normal, the statistical conclusion would have to be revised.","tokens_in":12914,"feed_emoji":"☀️","tokens_out":5734,"duration_ms":61296,"temperature":0.7,"pith_summary":"The paper sets out to show that a single-stage object detector built on YOLOv8, with three targeted modifications, can replace slow manual or image-processing identification of solar active regions. Trained on 26,531 labels from 6,975 SDO/HMI line-of-sight magnetograms spanning 2010 to 2019, and tested on temporally disjoint periods, the system reports an average precision of 94%, a precision of 0.89, and a recall of 0.88 while running in under one second per full-disk image. The same detector then feeds a statistical study in which the area and magnetic flux of active regions are both best described by log-normal distributions. If these results hold, researchers gain an automatic, fast way to build active-region catalogs for space-weather and solar-dynamo studies without waiting for manual annotation.","feed_headline":"One-second AI detector finds solar active regions at 94% precision","feed_subtitle":"A YOLOv8-based system reads solar magnetograms in under a second and maps likely flare source regions.","key_machinery":"The load-bearing object is the SARD network itself: YOLOv8s with deformable convolution (learnable offsets added to each sampling position, so the kernel can follow irregular active-region boundaries), an Efficient Channel Attention module (a fast 1D convolution that reweights feature channels based on global average pooling), and an extra detection head at 256x256 resolution for small regions. Supporting it is the dataset-generation pipeline: Otsu thresholding, morphological opening, region growing, morphological closing, then manual alignment of the resulting boxes to the NOAA catalog. The statistical claim is carried by fitting three candidate functions (log-normal, exponential, power law) to the area and flux histograms and testing the fits with a Kolmogorov-Smirnov test.","core_discovery":"On the paper's own terms, SARD is the discovery that a YOLOv8s base model extended with deformable convolution in its C2f blocks, an Efficient Channel Attention module, and an extra 256x256 small-object detection head outperforms every baseline tested: it reaches an average precision of 0.943, precision 0.89, recall 0.881, and F1 0.886 on the full test set assembled with a 14-day gap separating training and test images. Ablation experiments attribute AP gains of 0.2% to the extra head, 1.1% to ECA, and 0.9% to deformable convolution. When the detections are projected with a cylindrical equal-area projection and restricted to a central-meridian window, the area and magnetic flux distributions of 1,331 active regions are fit by log-normal curves, with Kolmogorov-Smirnov p-values above 0.05, while exponential and power-law fits are rejected. The paper reads this as support for the idea that magnetic flux elements fragment randomly in the convection zone before emerging as active regions.","pith_inferences":["One extension the paper leaves implicit: because low-activity periods trigger false positives on diffuse strong-field regions, a production version would likely need confidence calibration or a quiet-Sun suppression term before being trusted for flare-forecasting inputs.","A reader could test whether the log-normal result is an artifact of the NOAA-based bounding boxes, which draw human-defined boundaries around whole groups; an independent segmentation into individual magnetic concentrations might change the tail of the distribution.","The paper's comparison with YOLOv11 and YOLOv12 suggests newer object-detection backbones are not automatically better for this task; a natural follow-up is an architecture search over attention placement rather than scaling width and depth.","If SARD were retrained on vector magnetograms or on more recent data from solar cycle 25, the same pipeline would yield a near-real-time catalog that could be checked against flare occurrence to test whether detection confidence predicts eruption likelihood."],"forward_implications":["SARD can process a full-disk magnetogram in under one second, making it practical to run continuously on SDO/HMI data streams and to reprocess the entire 2010-2019 archive cheaply.","Because the test sets share no time window with the training set, the 94% average precision indicates the detector generalizes across the solar cycle rather than memorizing daily patterns.","The log-normal area and flux distributions imply that a fragmentation process, rather than a single characteristic scale, governs active-region emergence.","The ablation ordering shows that the small-object detection head, channel attention, and deformable convolution each contribute additively, so further architecture work along these lines is a plausible route to higher recall on small, clustered regions."],"supporting_citations":[{"why":"Supplies the traditional image-processing pipeline and baseline metrics (precision 0.778, recall 0.738) that SARD is compared against.","marker":"Zhang, Wang, and Liu (2010)"},{"why":"Source of the thresholding-morphology-region-growing labeling procedure adapted for building the training set.","marker":"Wang, Jiang, and Luo (2023)"},{"why":"Earlier YOLOv3 active-region detector whose 94% recall on a solar-minimum test set is the comparison the paper improves on in harder conditions.","marker":"Quan et al. (2021)"},{"why":"Establishes the log-normal sunspot-area distribution and the fragmentation interpretation that the paper's statistics confirm for active-region flux and area.","marker":"Bogdan et al. (1988)"},{"why":"Defines deformable convolution, the mechanism SARD uses to handle irregular active-region boundaries.","marker":"Dai et al. (2017)"},{"why":"Defines the ECA attention module used to emphasize small and clustered active regions.","marker":"Wang et al. (2020b)"},{"why":"Describes the HMI instrument on SDO that produced the 720-s line-of-sight magnetograms used as the dataset.","marker":"Scherrer et al. (2012)"},{"why":"Supplies the Otsu threshold method used in the first preprocessing step of the labeling pipeline.","marker":"Otsu et al. (1975)"}],"fun_headline_variants":["AI detects solar active regions in under a second with 94% precision","YOLOv8 spots sunspot regions in magnetograms at 94% AP","Log-normal flux emerges from AI-mapped solar active regions","Fast YOLOv8 detector maps solar flare regions with 94% precision","AI finds solar active regions, revealing log-normal size distributions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the NOAA active-region catalog, as reinterpreted through the paper's semi-automated labeling pipeline (thresholding, morphology, region growing, and manual adjustment), gives reliable and complete ground truth; if those labels are systematically biased, the reported 94% average precision and the log-normal statistics inherit that bias.","fun_headline_variants_meta":{"raw":{"variants":["AI detects solar active regions in under a second with 94% precision","YOLOv8 spots sunspot regions in magnetograms at 94% AP","Log-normal flux emerges from AI-mapped solar active regions","Fast YOLOv8 detector maps solar flare regions with 94% precision","AI finds solar active regions, revealing log-normal size distributions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000673,"raw_usage":{"total_tokens":3080,"prompt_tokens":979,"completion_tokens":2101,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":2018}},"tokens_in":595,"tokens_out":2101,"duration_ms":16725,"temperature":1.0,"reasoning_tokens":2018,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:16:06.439477+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build an independent test set of full-disk magnetograms in which active regions are manually outlined by several solar physicists according to a fixed magnetic-flux definition, deliberately excluding NOAA group boundaries and including weak and decayed regions; compute SARD's average precision at IoU>0.5 on that set. If it falls well below 0.94, the headline generalization fails; if the area and flux histograms from the independently segmented data are not log-normal, the statistical conclusion would have to be revised.","supporting_citations":[{"cited_title":"The Astrophysical Journal 723,","cited_arxiv_id":null,"evidence_quote":"Supplies the traditional image-processing pipeline and baseline metrics (precision 0.778, recall 0.738) that SARD is compared against."},{"cited_title":"Electronics 10,","cited_arxiv_id":null,"evidence_quote":"Earlier YOLOv3 active-region detector whose 94% recall on a solar-minimum test set is the comparison the paper improves on in harder conditions."},{"cited_title":"In: Proceedings of the IEEE international conference on computer vision,","cited_arxiv_id":null,"evidence_quote":"Defines deformable convolution, the mechanism SARD uses to handle irregular active-region boundaries."},{"cited_title":"Automatica 11,","cited_arxiv_id":null,"evidence_quote":"Supplies the Otsu threshold method used in the first preprocessing step of the labeling pipeline."}],"review_version":1}