{"id":"cd5a627d-f698-4882-9111-880422fce153","arxiv_id":"2508.18912","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A YOLOv11 variant with EfficientNet and squeeze-and-excitation attention is claimed to find PV thermal anomalies at 90.8% mAP, but no code, data, or baseline training details are provided.","lead":"HOTSPOT-YOLO is a drone-focused neural network that recombines a YOLOv11 detector with an EfficientNet backbone and attention blocks to locate hot spots in thermal images of solar panels. The paper reports a mean average precision of 90.8%, about 5.9 points above YOLOv11m, with a slightly smaller model.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline benchmark protocol is undocumented and internally inconsistent (dataset sizes 6000/720/450 vs 3600/470; Table 1 params off by ~1000x), so the 90.8% mAP and 5.9-point gain over YOLOv11m cannot be attributed to the architecture.","rationale":"The reader's weakest assumption identifies the uncontrolled baseline comparison as the key risk, which matches my read. The central claim is empirical and comparative; without a documented, reproducible protocol, the 90.8% mAP and the 5.9-point advantage cannot be verified or attributed to the model architecture. My stress-test reinforces this with the internal inconsistencies: two different dataset sizes (6000/720/450 in §2.3 vs 3600/470 in Table 2) and Table 1 parameter counts that are off by roughly three orders of magnitude from the stated convolution shapes. These inconsistencies mean even a motivated reader cannot reconstruct the experiment. The qualitative robustness tests and training curves are helpful but do not compensate for the missing code, weights, and baseline configurations. The verdict of REJECT is appropriate because the paper's primary claim is unsupported as written; no amount of architectural novelty can be assessed without a controlled benchmark. I agree with the reader, so no verdict adjustment is needed.","tokens_in":15534,"tokens_out":4372,"duration_ms":39027,"concrete_test":"Reconstruct the Roboflow thermal PV dataset, resolve the dataset-size contradiction, and train all Table 3 baselines plus HOTSPOT-YOLO with identical optimizer, epochs, batch size, augmentation, input resolution, and fixed seeds, evaluating on the same test split. If the mAP gap changes materially (e.g., YOLOv11m reaches ≥90% or HOTSPOT-YOLO falls below 84.9%), the central claim is unsupported. As a secondary check, recompute Table 1 parameter counts from the layer tensor shapes; if the total deviates >5% from 36.10M, the reported model complexity is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—HOTSPOT-YOLO's 90.8% mAP significantly outperforms baselines (Table 3, Figure 10)—requires that all models be trained and evaluated under identical conditions. The paper documents training only for HOTSPOT-YOLO (§2.3, Table 2). No baseline hyperparameters, augmentation, data splits, seeds, or preprocessing are given, and §2.3 states 6000/720/450 train/val/test images while Table 2 lists 3600/470. Additionally, Table 1's per-layer parameter counts are inconsistent with the stated tensor shapes: e.g., a 3x3x3→32 Conv2D has ~0.9K parameters, not 0.9M; the summed 36.10M total does not follow from the listed layers. The preprocessing row also shows 640x640x3 input resized to 224x224x3, conflicting with the 640x640 input used elsewhere. Because the experimental protocol cannot be reconstructed, the reported 5.9-point improvement may reflect differences in training setup (data split, augmentation, epochs, seeds) rather than the EfficientNet+SE architectural changes. This is a testability failure, not a judgment on intent.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HOTSPOT-YOLO, a YOLOv11 variant that replaces the backbone with EfficientNet and adds Squeeze-and-Excitation attention, targeting thermal anomaly detection in drone-based PV inspection. The authors report 90.8% mAP@0.5 on a dataset assembled from Roboflow, a 5.9-point improvement over YOLOv11m, and lower parameter/FLOP counts. The manuscript includes qualitative robustness tests under brightness/contrast, grayscale, blur, and ground-based imaging, and a benchmark against six detectors in Table 3. The central claim is that the proposed architecture is both more accurate and lighter than the compared baselines.","tokens_in":15886,"tokens_out":8044,"duration_ms":72278,"significance":"If the reported results were reproducible, the contribution would be practically useful: a lightweight, real-time thermal anomaly detector for PV inspections. The qualitative robustness demonstrations in Figures 7-9 are a genuine strength, and the SE/multi-scale aggregation formulation is clearly described. However, the paper's central numerical claims cannot be independently verified: the dataset is unnamed, code and weights are absent, baseline training protocols are not documented, and several internal inconsistencies affect the architecture and split sizes. No ablations support the attribution of the gains to the EfficientNet backbone and SE attention. As an empirical benchmark paper, these issues are load-bearing rather than cosmetic.","major_comments":[{"comment":"The dataset is not identifiable and the split sizes are contradictory. §2.3 says 'publicly available thermal imagery datasets sourced from Roboflow' but names no dataset, version, license, or class balance. It also states 6000 training / 720 validation / 450 test images, while Table 2 lists 3600 training and 470 validation images. Since every mAP number in the paper depends on this dataset, the discrepancy must be corrected and the dataset must be identified or released. No code or trained weights are provided either.","section":"§2.3 and Table 2"},{"comment":"The baseline benchmark protocol is undocumented. Table 3 reports mAP, parameters, FLOPs, and inference time for Cascade R-CNN, CenterNet, Faster R-CNN, YOLOv5m, YOLOv9m, and YOLOv11m, but the manuscript never states how these models were trained, on which split, with which hyperparameters, augmentations, data preprocessing, or seeds. Only HOTSPOT-YOLO's settings appear in Table 2. Without this information, the claimed 5.9-point improvement over YOLOv11m cannot be attributed to the architecture; it may reflect protocol differences rather than the proposed modifications.","section":"§3.6, Table 3"},{"comment":"The layer-wise architecture table is internally inconsistent. The first Conv2D layer (3x3, 3→32 channels) has 864 weights plus bias, i.e., ~0.0009M parameters, not the listed 0.9M. The depthwise/pointwise counts are similarly off by roughly 1000x, and the entries do not sum to the reported 36.10M. The preprocessing row also lists 224x224x3 output, contradicting the 640x640 input used throughout the paper. Consequently the claimed parameter reduction and 'lightweight' characterization are not supported by the provided specification.","section":"Table 1"},{"comment":"The headline result is reported on the validation split, not a held-out test set: §2.3 says the final mAP of 90.8% was obtained 'using the validation dataset.' Table 3 does not state whether the baseline numbers are validation or test mAP, and no confidence intervals or multiple seeds are given. This makes it impossible to judge whether the 5.9-point gap is meaningful or an artifact of validation-based model selection.","section":"§2.3 and §3.6"},{"comment":"The causal attribution to EfficientNet and SE attention is not supported by ablations. No experiment removes the SE blocks, swaps the backbone, or uses the original YOLOv11 backbone/neck on the same data. Given the parameter-count inconsistencies in Table 1, the architecture itself is not uniquely specified, so the statement in the abstract and conclusion that the improvements are due to these components is premature.","section":"§1, §3.6"}],"minor_comments":[{"comment":"The results section jumps from §3.4 to §3.6; §3.5 is missing.","section":"Section ordering"},{"comment":"Several entries appear to be transcription errors: Faster R-CNN has identical values for parameters and FLOPs (41.12), and YOLOv11m has identical values for FLOPs and time (29.61). Please check.","section":"Table 3"},{"comment":"In §3.6, reference [40] is cited twice in the same sentence; the second citation should likely be [41] or [42].","section":"References"},{"comment":"The caption/text says the final mAP of 90.8% was reached after 120 epochs, while Table 2 says training ran for 200 epochs with early stopping disabled. Clarify the relationship between these statements.","section":"Figure 3"},{"comment":"There is a typo: 'Iamge' should be 'Image'.","section":"Appendix D"}],"recommendation":"reject","confidential_remarks":"The editor should weigh that the manuscript's central numerical claims are not checkable as submitted: the dataset is unnamed, code/weights are absent, baseline training details are missing, and Table 1's parameter counts are inconsistent with the stated tensor shapes. These are not simple presentation issues; they undermine the empirical contribution. If the authors can provide a complete reproducibility package and correct the architecture table, a resubmission might be considered, but on the current manuscript I cannot recommend acceptance or minor revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here’s my take. The paper is a straightforward engineering combination: YOLOv11 frame, EfficientNet backbone, SE attention. That’s a legitimate incremental step, and the qualitative results—especially the robustness checks on brightness, blur, and ground-based images—show the author is thinking about deployment. I believe the numbers when they are consistent with what is shown.\n\nThe problem is that the central number, the 90.8% mAP, cannot be verified. There is no code, no weights, no dataset name, and no description of how the baselines were trained. The paper’s own tables contradict the text: Section 2.3 says 6000/720/450 train/val/test, Table 2 says 3600/470. The layer-table parameter counts are off by roughly a factor of 1000 for a 3x3x3->32 convolution, and the stated total (36.10M) doesn’t follow from the layers listed. The preprocessing row says 224x224 while the rest of the paper uses 640x640. These aren’t cosmetic typos; they mean the experimental protocol cannot be reconstructed.\n\nThat is a testability failure, not evidence of fraud. The paper honestly notes its ground-based limitation, and it cites the relevant prior work on EfficientNet and SE for PV defects. I’d give credit for that.\n\nBut as a reader, I would not trust the benchmark comparison until the author releases the dataset (or at least its name), the training/evaluation protocol, and the code. Without that, the 5.9-point improvement over YOLOv11m could just be different hyperparameters or data splits.\n\nWould I bring it to a reading group? Maybe, as an example of why reporting standards matter. Would I cite it? No. Would I send it to peer review? No, not in this form—an editor should desk-reject until the evidence is available. The author has a plausible method; this needs a major revision, not a referee report.","headline":"A plausible architecture and some nice qualitative demos, but the central mAP claim is unverifiable because the benchmark protocol is missing and the tables contradict the text; not ready for review.","tokens_in":16300,"tokens_out":4361,"would_cite":false,"duration_ms":38697,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that HOTSPOT-YOLO—a YOLOv11 variant with an EfficientNet backbone and Squeeze-and-Excitation attention—detects thermal anomalies in drone-based PV images at 90.8% mean average precision, beating YOLOv11m by 5.9 points with","keywords":["thermal anomaly detection","solar photovoltaic inspections","HOTSPOT-YOLO","YOLOv11","EfficientNet","Squeeze-and-Excitation attention","drone thermography","mean average precision"],"falsifier":"Train all six comparison detectors and HOTSPOT-YOLO on the same public thermal PV image set with identical splits, augmentation, optimizer settings, epochs, and random seeds. If YOLOv11m reaches or exceeds 90.8% mAP, or if HOTSPOT-YOLO falls to the 84.9% level, the claimed architectural advantage collapses and the difference would instead reflect protocol differences.","tokens_in":15440,"feed_emoji":"🔥","tokens_out":10426,"duration_ms":95095,"temperature":0.7,"pith_summary":"Solar panels develop small hotspots that drone thermal cameras can catch only if an automated detector is both accurate enough to see subtle temperature differences and light enough to run in real time. HOTSPOT-YOLO is the paper's attempt to build that detector by grafting an EfficientNet backbone and Squeeze-and-Excitation attention onto YOLOv11. The paper reports 90.8% mean average precision on thermal PV images, a 5.9-point gain over YOLOv11m, with 2.12 million fewer parameters and real-time inference. The practical stake is straightforward: if the result holds, routine drone sweeps of utility-scale solar plants could flag faulty modules sooner and run on cheaper onboard hardware.","feed_headline":"A lightweight YOLO finds solar hotspots at 90.8% precision","feed_subtitle":"Thermal-drone model beats YOLOv11 by 5.9 points with 2.12 million fewer parameters.","key_machinery":"The load-bearing mechanism is channel recalibration: Squeeze-and-Excitation blocks compress each feature map to a per-channel scalar via global average pooling, learn two fully connected layers that assign importance weights, and multiply those weights back into the feature maps. That is what lets the model emphasize thermally significant channels while suppressing background. Around this core, EfficientNet's depthwise separable convolutions keep the parameter and FLOP counts low, and the feature-aggregation block adds low-, mid-, and high-level feature maps so small hotspots and larger defective regions are seen by the same detection head.","core_discovery":"On its own terms, the paper advances one architectural recipe: keep the YOLOv11 detection head, replace feature extraction with depthwise-separable EfficientNet blocks, interleave SE blocks that recalibrate feature channels after global average pooling, and fuse low-, mid-, and high-level feature maps before prediction. Trained for 200 epochs with Adam and cosine learning-rate decay on a public collection of thermal PV images, the model is reported to reach 90.8% mAP@0.5, beat YOLOv11m by 5.9 points, and run at 25.22 ms per image with 36.10M parameters and 25.53 GFLOPs. The paper also reports that confidence falls under heavy blur and for ground-based oblique images, which it explains as a c","pith_inferences":["A natural next step the paper does not report is a controlled re-run of all baselines under the exact same training protocol; that experiment alone would separate the architecture's contribution from setup differences.","The EfficientNet-plus-SE recipe is not solar-specific; it should transfer to other small-anomaly thermal domains, and can be tested directly on public thermal fault datasets for power equipment or buildings.","Appendix A's center-weighted bounding-box distribution suggests a per-location stratified mAP would reveal whether accuracy holds across the full drone frame.","Since blur clearly lowers confidence, blur augmentation during training is a direct, testable fix that might close that gap."],"forward_implications":["At 90.8% mAP, automated aerial inspection would catch a larger share of small hotspot defects than the 84.9% YOLOv11m baseline, reducing missed faults during routine surveys.","At 36.10M parameters and 25.53 GFLOPs, the model fits the compute envelope of drone-class hardware while still running at 25.22 ms per frame.","The reported stability under brightness, contrast, grayscale, and blur transformations means a single deployed model could handle variable weather and camera settings without per-site retraining.","The weaker results on ground-based oblique images imply the model works best for drone flight paths as trained, and that expansion to handheld inspections would need additional fine-tuning data."],"supporting_citations":[{"why":"Defines the YOLOv11 architecture that HOTSPOT-YOLO starts from and whose detection head and loss are retained.","marker":"[18]"},{"why":"Supports using an EfficientNet backbone for high-noise solar panel defect identification.","marker":"[19]"},{"why":"Supplies the Squeeze-and-Excitation attention mechanism for PV hotspot detection.","marker":"[20]"},{"why":"Motivates the multi-level context aggregation step used to fuse low-, mid-, and high-level thermal features.","marker":"[22]"},{"why":"Provides the Cascade RCNN baseline in the benchmark table that the proposed model claims to outperform.","marker":"[34]"},{"why":"Provides the CenterNet baseline in the benchmark comparison.","marker":"[35]"},{"why":"Provides the Faster RCNN baseline in the benchmark comparison.","marker":"[36]"},{"why":"Provides the YOLOv5m baseline in the benchmark comparison.","marker":"[37]"},{"why":"Provides the YOLOv9m baseline in the benchmark comparison.","marker":"[38]"},{"why":"Provides the YOLOv11m baseline, the closest comparison whose 84.9% mAP and 38.22M parameters are central to the claimed improvement.","marker":"[39]"}],"fun_headline_variants":["New YOLO model pulls ahead by 5.9 points on solar hotspots","Drone thermal AI: YOLO variant beats original by 5.9 mAP","Tiny YOLO for solar panels: 90.8% precision, 5.9-point gain","HOTSPOT-YOLO: light thermal drone AI hits 90.8% mAP"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The benchmark comparison assumes every baseline was trained and evaluated under exactly the same data splits, preprocessing, augmentation, and hyperparameters as HOTSPOT-YOLO, but the paper specifies training details only for HOTSPOT-YOLO.","fun_headline_variants_meta":{"raw":{"variants":["New YOLO model pulls ahead by 5.9 points on solar hotspots","Drone thermal AI: YOLO variant beats original by 5.9 mAP","Tiny YOLO for solar panels: 90.8% precision, 5.9-point gain","HOTSPOT-YOLO: light thermal drone AI hits 90.8% mAP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001447,"raw_usage":{"total_tokens":5655,"prompt_tokens":724,"completion_tokens":4931,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":468,"completion_tokens_details":{"reasoning_tokens":4844}},"tokens_in":468,"tokens_out":4931,"duration_ms":34529,"temperature":1.0,"reasoning_tokens":4844,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:05:21.979317+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train all six comparison detectors and HOTSPOT-YOLO on the same public thermal PV image set with identical splits, augmentation, optimizer settings, epochs, and random seeds. If YOLOv11m reaches or exceeds 90.8% mAP, or if HOTSPOT-YOLO falls to the 84.9% level, the claimed architectural advantage collapses and the difference would instead reflect protocol differences.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports using an EfficientNet backbone for high-noise solar panel defect identification."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Squeeze-and-Excitation attention mechanism for PV hotspot detection."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates the multi-level context aggregation step used to fuse low-, mid-, and high-level thermal features."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Cascade RCNN baseline in the benchmark table that the proposed model claims to outperform."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the CenterNet baseline in the benchmark comparison."},{"cited_title":"(2023, May)","cited_arxiv_id":null,"evidence_quote":"Provides the Faster RCNN baseline in the benchmark comparison."},{"cited_title":"Link: https://docs.ultralytics.com/models/yolov5/#overview Page 29 of 29","cited_arxiv_id":null,"evidence_quote":"Provides the YOLOv5m baseline in the benchmark comparison."},{"cited_title":"Link: https://docs.ultralytics.com/models/yolov9/#what-tasks-and- modes-does-yolov9-support","cited_arxiv_id":null,"evidence_quote":"Provides the YOLOv9m baseline in the benchmark comparison."},{"cited_title":"Link: https://docs.ultralytics.com/models/yolo11/","cited_arxiv_id":null,"evidence_quote":"Provides the YOLOv11m baseline, the closest comparison whose 84.9% mAP and 38.22M parameters are central to the claimed improvement."}],"review_version":1}