{"id":"a3c696db-7c9b-44a3-9ff6-be3ec4551180","arxiv_id":"2509.08026","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":1.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"SI-EDTL, a 2022 ensemble of three pre-trained Faster R-CNNs with five classifiers tuned by WOA, is re-presented here with 91.3% accuracy on AU-AIR, but no code, error bars, or fair baseline comparisons are provided.","lead":"This paper re-publishes a shortened version of the authors' 2022 SI-EDTL model, an ensemble of three Faster R-CNN extractors and five classifiers tuned by whale optimization, reporting 91.3% accuracy on the AU-AIR UAV dataset. It offers no new experimental evidence beyond the 2022 paper and compares against baselines that are not optimized, so the claimed superiority is not well supported.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2 is the only support for the claimed superiority, but its baselines are unspecified and explicitly unoptimized, so the performance gap is not currently evidenced.","rationale":"The preprint opens by stating that it is a shortened version of a 2022 CCPE article, and its figures and tables are cited as [48], so the empirical content is not new to this submission. Even setting novelty aside, the correctness of the central performance claim depends entirely on the Table 2 comparison. The text itself discloses that the mobile detectors were not optimized, and the implementation details for all baselines are absent. This is not a disagreement with current consensus but a missing-evidence problem: no current information rules out the possibility that the reported gap is due to baseline under-configuration or a mismatched evaluation protocol. A concrete reproduction of the two mobile detectors under a fair protocol with moderate tuning would settle whether the gap is real. I therefore agree with the reader's weakest-assumption analysis and find no reason to change the REJECT verdict.","tokens_in":10591,"tokens_out":6988,"duration_ms":61864,"concrete_test":"Re-run Table 2 using open-source YOLOv3-Tiny and MobileNetv2-SSDLite implementations on the exact AU-AIR train/test region split of Table 1, with a modest hyperparameter search (learning rate, epochs, anchor scales, input size) and identical accuracy/precision/recall computation; if either tuned baseline approaches or exceeds the SI-EDTL row, or the gap shrinks substantially, the claimed superiority is a comparison artifact rather than a method advantage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The preprint's central claim—that SI-EDTL's 91.3% accuracy, 89.3% precision, and 89.1% recall constitute superior detection—is supported only by Table 2. For that comparison to be probative, every baseline must be trained and evaluated under the same protocol on the same AU-AIR region split and scored with the same metric. Section 4.2 does not state how YOLOv3-Tiny or MobileNetv2-SSDLite were implemented (framework, anchors, input resolution, epochs, augmentation, or whether they operated on full images or cropped regions). The text itself concedes that the low baseline scores are explained by 'not optimizing the default parameters for aerial images,' which means the gap may be a tuning artifact rather than a method advantage. No error bars, per-class breakdown, confidence intervals, or code are supplied; with only 23 test Bus samples and a heavily imbalanced Table 1, single-run aggregate accuracy is unstable. The conclusion further claims superiority in mean average precision, but no mAP is reported anywhere. Because the only empirical evidence for superiority is this under-specified comparison, the central claim is currently unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes SI-EDTL, a two-stage ensemble deep transfer learning model for multi-vehicle detection in UAV imagery. The first stage uses three pre-trained Faster R-CNN feature extractors (InceptionV3, ResNet50, GoogLeNet), and the second stage uses five classifiers (KNN, SVM, MLP, C4.5, Naive Bayes), yielding 15 base learners whose weighted outputs are aggregated and thresholded. The ensemble weights and decision threshold are tuned on the training set with the Whale Optimization Algorithm using a fitness that combines accuracy, precision, and recall. The paper reports experiments on the AU-AIR dataset, claiming 91.3% accuracy, 89.3% precision, and 89.1% recall, and states that this outperforms existing detectors including YOLOv3-Tiny and MobileNetv2-SSDLite.","tokens_in":10970,"tokens_out":5945,"duration_ms":56351,"significance":"If the empirical claims were rigorously supported, the two-stage ensemble idea and the use of WOA to tune ensemble weights and threshold would be a plausible engineering contribution to UAV-based vehicle detection. The architecture is described in enough detail to be reproduced in principle, and the explicit fitness formulation in Eqs. (4)-(7) is a useful starting point. However, the paper's central claim of superiority rests entirely on a single comparison table with under-specified baselines, no error bars, no statistical tests, and no reported detection metric such as mAP, despite the conclusion asserting an mAP advantage. In its current form the manuscript provides no machine-checked proofs, no released code, and no reproducible evaluation protocol, so the claimed performance gain is not evidenced.","major_comments":[{"comment":"The central claim that SI-EDTL outperforms existing methods is not supported because the baseline comparisons are not established under the same protocol. The manuscript gives no implementation details for YOLOv3-Tiny or MobileNetv2-SSDLite: no framework, input resolution, anchors, epochs, augmentation, optimizer, or training/evaluation split. It also does not state whether these detectors operated on full images, cropped regions, or the same region proposals as SI-EDTL. The text itself attributes their low scores to \"not optimizing the default parameters for aerial images,\" which concedes that the gap may be a tuning artifact rather than a method advantage.","section":"Section 4.2, Table 2"},{"comment":"The evaluation protocol is ambiguous and appears to mix different tasks. Training is described as using cropped target objects and background regions, and Eqs. (5)-(7) define region-level classification metrics, not detection metrics. Yet Table 2 compares against full-image detectors such as YOLOv3-Tiny. It must be stated explicitly whether all methods are evaluated on the same set of region proposals with the same overlap/decision criterion; otherwise the accuracy, precision, and recall values in Table 2 are not comparable quantities.","section":"Sections 3.2 and 4.2"},{"comment":"No variance, confidence intervals, or statistical significance tests are reported anywhere. Table 1 shows that the test set contains only 23 Bus samples, and the class distribution is heavily imbalanced, so a single-run aggregate accuracy is unstable. The authors should report per-class results, repeated runs or cross-validation splits, and appropriate significance tests before claiming superiority.","section":"Section 4.2, Table 2"},{"comment":"The conclusion states that SI-EDTL outperforms other methods \"in accuracy, precision, recall, and mean average precision,\" but no mAP value is reported in Section 4 or in Table 2. Either provide the mAP evaluation with the same protocol as the other methods, or remove the mAP claim from the conclusion.","section":"Section 5"}],"minor_comments":[{"comment":"The header 'TRAN DATASET' should be corrected to 'TRAIN DATASET'.","section":"Table 1"},{"comment":"The heading 'Comparision with existing methods' contains a typo; it should read 'Comparison with existing methods'.","section":"Section 4.2 heading"},{"comment":"The first sentence contains a typo: 'de ep' should be 'deep'.","section":"Section 5"},{"comment":"The symbols N1Train and N0Train are introduced but never quantified; please state the actual number of cropped object and background training samples used.","section":"Section 3.2"},{"comment":"All figure and table captions cite the authors' earlier publication [48] as the source. Because this manuscript is explicitly a shortened version of that paper, the authors should clarify what new content is added here and ensure that any reused material is properly credited or reproduced with permission.","section":"Figures and tables"}],"recommendation":"reject","confidential_remarks":"The manuscript is explicitly a shortened version of a 2022 paper in Concurrency and Computation: Practice and Experience (reference [48]). If this is intended as a new submission, the lack of incremental novelty and the near-duplicate content are serious editorial concerns. Even setting that aside, the empirical evidence is insufficient: the baseline comparison is under-specified and unfair, and the claimed mAP superiority is not reported. The technical problems are load-bearing and would require substantial new experiments and a reworked evaluation to address."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper is honest about being a shortened version of the authors' 2022 CCPE article, so it contains no new result and does not claim one. Second, the only empirical evidence for the method's superiority is Table 2, and that table is not sufficiently specified to carry the claim.\n\nWhat is genuinely useful here is the clarity of the description. The two-stage ensemble (three pre-trained Faster R-CNN feature extractors, five classifiers, WOA weight/threshold tuning) is explained in enough detail to reproduce the architecture. The authors also explicitly state the relationship to the earlier article and reuse figures with attribution to [48]. That is more transparent than many duplicate postings.\n\nThe soft spots are real but mostly boil down to one thing. The comparison in Table 2 lacks implementation details for the baselines. YOLOv3-Tiny and MobileNetv2-SSDLite are reported at 46.1% and 55.3% accuracy, which is far below what those detectors typically achieve; the text explains this by saying default parameters were not optimized for aerial images. That may be true, but it means the gap is a tuning artifact as much as a method advantage. There are no error bars or confidence intervals, and the Bus class has 23 test samples, so the aggregate numbers are unstable. The conclusion also claims superior mean average precision, but no mAP is reported anywhere. These are not fatal to the architecture, but they are fatal to the claim that this preprint demonstrates superiority.\n\nWho is this for? Someone who wants a compact statement of the SI-EDTL method and its published numbers, or a reviewer checking whether the arXiv text matches the 2022 article. For a new-result audience, there is nothing to learn. I would not cite the preprint version when the journal article exists. I would not bring it to reading group. If this were submitted as a new contribution, I would desk-reject because it duplicates already published work; if it is meant as a self-archived copy, that is fine, but it should be labeled as such and not marketed as a fresh result.","headline":"A transparent self-archive of a 2022 journal article, with a useful method description but an under-specified comparison table that cannot support the claimed superiority.","tokens_in":11391,"tokens_out":1863,"would_cite":false,"duration_ms":16661,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a two-stage ensemble of 15 transfer-learned detectors—three Faster R-CNN feature extractors and five classifiers—with whale-optimized weights achieves 91.3% accuracy on vehicle detection in UAV images.","keywords":["Deep learning","Transfer learning","Ensemble learning","Multiple vehicle detection","UAVs","Faster R-CNN","Whale optimization algorithm","AU-AIR dataset"],"falsifier":"Re-run YOLOv3-Tiny and MobileNetv2-SSDLite on the same AU-AIR training and test split with their own hyperparameters tuned on a validation fold, using the same metric definitions as SI-EDTL; if either reaches or exceeds 91.3 percent accuracy, the claimed superiority is not supported.","tokens_in":10415,"feed_emoji":"🚁","tokens_out":9105,"duration_ms":69287,"temperature":0.7,"pith_summary":"This paper sets out to show that vehicle detection in UAV aerial imagery can be improved by combining many transfer-learned components instead of training a single large detector. It combines three pre-trained Faster R-CNN feature extractors (InceptionV3, ResNet50, GoogLeNet) with five classifiers (KNN, SVM, MLP, C4.5, Naive Bayes), forming 15 base learners whose class votes are combined by weighted averaging. The ensemble weights and a decision threshold are tuned with the whale optimization algorithm rather than set by hand. On the AU-AIR benchmark the model reports 91.3 percent accuracy, 89.3 percent precision, and 89.1 percent recall, exceeding the paper's listed baselines. If that comparison is fair, this makes a tunable ensemble of existing models a practical route for aerial traffic monitoring without training a deep detector from scratch.","feed_headline":"Two-stage ensemble detects vehicles in UAV images at 91.3%","feed_subtitle":"Three feature extractors times five classifiers, fused by whale-tuned weights, lifts accuracy past single models.","key_machinery":"The load-bearing object is the two-level ensemble of $N_{FE}\\times N_{CL}=15$ base learners with weighted-average fusion. Each base learner is one classifier from {KNN, SVM, MLP, C4.5, Naive Bayes} applied to features from one of three Faster R-CNN feature extractors (InceptionV3, ResNet50, GoogLeNet). The fused score decides the region's class only if it exceeds a decision threshold $D_{Th}$; both the 15 ensemble weights and $D_{Th}$ are the hyperparameters optimized by the whale optimization algorithm (a swarm search that mimics humpback-whale hunting) against a fitness of 0.5 accuracy plus 0.3 average precision plus 0.2 average recall. This mechanism lets the system combine complementary strengths of different architectures and classifiers without training a new deep network, and it makes the accuracy-precision-recall trade-off tunable after training.","core_discovery":"The central claim is that a two-stage ensemble of deep transfer learners, called SI-EDTL, detects Cars, Vans, Trucks, and Buses in UAV images more accurately than single detectors. In the first stage, three ImageNet-pretrained CNNs are converted into Faster R-CNN feature extractors by attaching a region proposal network, RoI pooling, and a bounding-box regression layer to chosen intermediate features ('mixed7', 'activation40_relu', 'inception_4d-output'). In the second stage, each extractor feeds five classifiers, giving 15 base learners. For each region proposal, the final class score is the weighted average of the learners' binary votes, and the proposal is assigned to the class with the highest score if that score exceeds a tuned threshold. The whale optimization algorithm sets the 15 weights and the threshold using a fitness that balances accuracy, precision, and recall under 10-fold cross-validation, and the resulting system reports 91.3 percent accuracy, 89.3 percent precision, and 89.1 percent recall on the AU-AIR test split.","pith_inferences":["The paper reports only whole-dataset accuracy, precision, and recall; a per-class breakdown, especially for Bus with only 23 test samples, would show whether the ensemble's gain is consistent across rare classes or driven by Car and Truck.","Because the method is modular, an ablation dropping one feature extractor or one classifier would isolate which components carry the gain; the paper does not include such an ablation.","The same weighted-voting recipe could be applied to other region-proposal detectors and other aerial datasets, but because only AU-AIR is tested, the method's generality is an untested inference rather than a paper claim."],"forward_implications":["If the reported results hold, SI-EDTL beats all six listed baselines on AU-AIR, including the strongest prior deep method (SW-CNN at 83.5 percent accuracy) and both mobile detectors, on the same dataset split.","The ensemble adds little online cost: because the three Faster R-CNN extractors run on parallel GPUs, per-image test time is 1.57 seconds, close to the slowest single extractor (InceptionV3 at 1.43 seconds).","Offline training time is dominated by the three Faster R-CNN extractors (about 19.9 hours total), while the whale-optimization tuning stage is comparatively short, so the method is a viable offline-train and online-deploy pipeline.","The tunable fitness weights mean an operator can rebalance precision versus recall by rerunning whale optimization, without retraining the 15 base learners."],"supporting_citations":[{"why":"provides the Faster R-CNN framework whose pre-trained CNNs are converted into feature extractors","marker":"[27]"},{"why":"supplies the whale optimization algorithm used to tune the 15 ensemble weights and the decision threshold","marker":"[57]"},{"why":"is the AU-AIR dataset on which the model is trained and evaluated","marker":"[59]"},{"why":"is the original SI-EDTL article from which this shortened version and its figures and tables are taken","marker":"[48]"},{"why":"is the YOLOv3-Tiny baseline that SI-EDTL must beat in Table 2","marker":"[63]"},{"why":"is the MobileNetv2-SSDLite baseline that SI-EDTL must beat in Table 2","marker":"[64]"},{"why":"is the SW-CNN deep baseline, the strongest prior method in the comparison table","marker":"[35]"},{"why":"is the HOG+SVM baseline used as a classical machine-learning comparison","marker":"[33]"}],"fun_headline_variants":["Swarm-optimized ensemble detects UAV vehicles at 91.3%","Whale-tuned deep learners hit 91.3% in aerial vehicle detection","Two-stage ensemble: 91.3% accuracy for UAV vehicle detection","Fusing 15 classifiers with swarm intelligence for UAV detection","Deep transfer ensemble boosts UAV vehicle detection to 91.3%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the baselines in Table 2 were run under a fair, comparable protocol; the paper does not describe how YOLOv3-Tiny and MobileNetv2-SSDLite were configured, so part of the performance gap could come from tuning or evaluation differences rather than from the ensemble itself.","fun_headline_variants_meta":{"raw":{"variants":["Swarm-optimized ensemble detects UAV vehicles at 91.3%","Whale-tuned deep learners hit 91.3% in aerial vehicle detection","Two-stage ensemble: 91.3% accuracy for UAV vehicle detection","Fusing 15 classifiers with swarm intelligence for UAV detection","Deep transfer ensemble boosts UAV vehicle detection to 91.3%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000714,"raw_usage":{"total_tokens":3188,"prompt_tokens":901,"completion_tokens":2287,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":2192}},"tokens_in":517,"tokens_out":2287,"duration_ms":14996,"temperature":1.0,"reasoning_tokens":2192,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:10:29.986008+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run YOLOv3-Tiny and MobileNetv2-SSDLite on the same AU-AIR training and test split with their own hyperparameters tuned on a validation fold, using the same metric definitions as SI-EDTL; if either reaches or exceeds 91.3 percent accuracy, the claimed superiority is not supported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the Faster R-CNN framework whose pre-trained CNNs are converted into feature extractors"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the whale optimization algorithm used to tune the 15 ensemble weights and the decision threshold"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"is the original SI-EDTL article from which this shortened version and its figures and tables are taken"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"is the SW-CNN deep baseline, the strongest prior method in the comparison table"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"is the HOG+SVM baseline used as a classical machine-learning comparison"}],"review_version":2}