{"id":"fe67c56f-0029-4721-8d57-a542d364f1b3","arxiv_id":"2411.17897","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Combining ResNet50 image features with a support vector machine predicted grapevine leaf area index from drone RGB images with MAE 0.32 and MSE 0.21, beating handcrafted feature methods in a 1,469-image comparison.","lead":"This paper reports that drone photos of single grapevine plants, combined with a pre-trained ResNet50 feature extractor and a support vector machine, estimate leaf area index with MAE 0.32 and MSE 0.21. The result is a faster, non-destructive alternative to hand measurements, but the paper omits key validation details such as train/test splitting.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table I's reported MSE/MAE are not demonstrably out-of-sample: no train/test split or cross-validation is described, and 1,469 crops extracted from 498 source images create group-leakage risk, so the ResNet+SVM superiority claim rests on an unverified evaluation protocol.","rationale":"The paper has positive aspects: the dataset is relevant, the comparison across feature extraction methods is systematic, and the use of a pre-trained ResNet50 with standard machine learning models is plausible and reproducible in principle. The direction of the reported result is consistent with findings in similar crops, so the concern is not that the approach is impossible or internally inconsistent. However, the numerical superiority claim depends entirely on whether the reported errors are out-of-sample. The reader identified this as the weakest assumption, and I agree. The absence of any documented evaluation protocol is not a cosmetic omission: high-dimensional ResNet features combined with a flexible RBF-kernel SVM can easily memorize a small, structured dataset, and crops from the same source image violate the independence assumption needed for a random crop-level split. The group-leakage issue is particularly important because 1,469 crops originate from only 498 drone images, so the effective number of independent samples may be much smaller than it appears. The concrete test of grouped cross-validation would settle whether ResNet+SVM retains its edge and whether the MSE remains near 0.21. Until that evidence is provided, the reader's CONDITIONAL verdict is appropriate; my read does not move it.","tokens_in":5888,"tokens_out":3030,"duration_ms":29117,"concrete_test":"Re-run the full nine-model comparison with 5-fold cross-validation grouped by the 498 source drone images (all crops from one source image stay in one fold), fixing all hyperparameters and thresholds to values chosen from training folds only and recording a fixed random seed. If ResNet+SVM's grouped out-of-fold MSE is not below the runner-up by a margin larger than the cross-validation spread, or if the MSE moves substantially above 0.21, the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ResNet50 features plus RBF-SVM gives LAI predictions with MSE 0.21 and MAE 0.32, and that this proves the superiority of deep features (Section III, Table I). For this claim to hold, the errors must be predictive errors on data not used to fit or tune the model. The paper never states a train/test split, cross-validation scheme, or random seed (Sections II-G and III). In addition, the 1,469 plant images are crops from only 498 drone images (Section II-A), so crops from the same source share illumination, background, and canopy context; a crop-level split leaks information and gives an optimistic bias. The Green Area pipeline reports thresholds that are 'consistent and correspond to the best results' (Section II-C), suggesting parameters may have been selected on the same data. Thus the reported MSE/MAE may be in-sample fit quality, and the claimed advantage of ResNet+SVM over simpler features could be an overfitting artifact rather than a predictive result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper compares three feature extraction pipelines (green-area edge detection, a color/shape/texture vocabulary, and pre-trained ResNet50 embeddings) combined with three regressors (linear regression, RBF support vector regression, and random forest) for estimating grapevine leaf area index from UAV RGB images. LAI labels are obtained from a SunScan canopy analysis system, and the dataset comprises 1,469 individual plant images extracted from 498 annotated drone images. The central empirical claim, based on Table I, is that ResNet50 features with SVM yield the lowest error (MSE 0.21, MAE 0.32, MAPE 34%) and that deep-learning features are generally superior to traditional features for this task.","tokens_in":6191,"tokens_out":3542,"duration_ms":30036,"significance":"If the reported errors are genuine out-of-sample predictive errors, the paper provides a practically useful, low-cost pipeline for non-destructive LAI estimation and a systematic comparison of feature extractors. Strengths include the use of real field data with external LAI measurements, explicit feature definitions, and multiple model/feature combinations as baselines. However, the significance of the central comparison depends entirely on the evaluation protocol, which is not described, so the result is currently an unverified performance claim rather than a demonstrated predictive advantage.","major_comments":[{"comment":"The paper reports only point estimates of MSE, MAE, and MAPE and never states whether the models were evaluated on holdout data. Sections II-G and III do not describe a train/test split, cross-validation scheme, or random seed. Since all hyperparameters (e.g., SVM RBF parameters, RF with 100 trees) and the ResNet embeddings are used on the same 1,469 images, the reported values may be in-sample fit statistics. Please specify the evaluation protocol, use repeated cross-validation (or a held-out set), and report confidence intervals for the metrics.","section":"Section III, Table I"},{"comment":"The 1,469 plant images are crops extracted from only 498 source drone images, so samples sharing a source image are not independent: they share illumination, background, and canopy context. A random split at the level of individual crops can leak information across training and test sets and inflate the apparent accuracy. The evaluation should be group-aware, e.g., by always keeping all crops from the same source image in the same fold, or by testing on a separate flight/field/trial.","section":"Section II-A"},{"comment":"The text states that the threshold values (pixel value 50, Canny 50/150) 'correspond to the best results.' If these thresholds were selected by examining performance on the full dataset, the Green Area features are tuned on the test data. Please clarify how the thresholds were chosen and, if they were tuned, use an inner validation loop or nested cross-validation so that model comparison in Table I is not biased by feature-level tuning.","section":"Section II-C"}],"minor_comments":[{"comment":"Fix typographical errors such as 'UA V' in the abstract, 'V ocabulary' in Table I, 'leafs' in Section III, and 'proofs' in Section III ('proves').","section":"Throughout"},{"comment":"The phrase 'The reported values for all images are consistent and correspond to the best results' is ambiguous; state which values are consistent and how 'best results' was determined.","section":"Section II-C"},{"comment":"Provide additional details on the SunScan measurement protocol (e.g., sampling scheme, number of below-canopy readings per plant) so readers can judge the quality of the LAI reference labels.","section":"Section II-A"},{"comment":"Figure 3 is referenced in the text but its content (pipeline) is not described in detail; a short description of the arrows and stages would help reproducibility.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"This is a straightforward empirical comparison; the main gap is methodological reporting rather than conceptual novelty. If the authors add a rigorous evaluation protocol with group-aware splits and uncertainty estimates, the paper could be publishable. The scope fits an applied remote sensing / precision agriculture venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful, modest empirical study — a new grapevine dataset with ground-truth LAI from a SunScan, and a clean comparison of three feature extractors (green-area, handcrafted, ResNet50) against three regressors (LR, SVM, RF). The finding that ResNet features plus SVR gives the lowest errors is plausible and consistent with similar studies in rice and wheat. The authors deserve credit for collecting real drone data and measuring LAI with an external instrument.\n\nThe soft spot is the evaluation protocol, and it is load-bearing. Section III reports only point estimates of MSE/MAE/MAPE with no train/test split, no cross-validation, and no random seed. Since the 1,469 plant images are crops from 498 source images, a random crop-level split leaks information through shared illumination and background, so the errors could be optimistic. The green-area thresholds are described as 'consistent and correspond to the best results' — that sounds like tuning on the same data. Because of this, the claimed advantage of ResNet+SVM (MSE 0.21, MAE 0.32) is not demonstrably out-of-sample, and the phrase 'proofs a strong fit' overclaims.\n\nThat said, the problems are fixable. The authors need to specify the split, cross-validate at the source-image level (or cluster crops by source image), report confidence intervals, and compare against a simple vegetation-index baseline like NDVI. I would also ask for a statistical test on the ranking of methods.\n\nThis paper is for practitioners in precision viticulture and for researchers looking for a baseline comparison on a new crop. It is not a methodological advance, but it is a legitimate empirical data point. I would send it to review — the subject is relevant and the data collection is real — but the revision needs to address the protocol before the numbers are cited.","headline":"Modest, useful grapevine LAI dataset and feature comparison, but the missing train/test protocol makes the headline numbers not yet trustworthy.","tokens_in":6687,"tokens_out":2834,"would_cite":false,"duration_ms":23349,"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 ResNet-plus-SVM pipeline estimates grapevine leaf area from drone RGB images with the lowest error of the nine combinations tested.","keywords":["Leaf Area Index","grapevine","UAV imagery","machine learning regression","deep learning feature extraction","ResNet50","support vector regression","precision agriculture"],"falsifier":"Re-run all nine feature-extractor/model combinations with a documented group-wise split that keeps every crop extracted from the same drone image inside the same training or test fold, and compare ResNet-plus-SVM against green-area-plus-random-forest; if the error gap closes or reverses, the claimed superiority of the deep-learning pipeline is an artifact of information leakage rather than genuine predictive quality.","tokens_in":5647,"feed_emoji":"🍇","tokens_out":5626,"duration_ms":46341,"temperature":0.7,"pith_summary":"This paper tries to establish that grapevine leaf area index (LAI) can be estimated automatically from ordinary drone RGB images by pairing a deep-learning feature extractor with a machine-learning regressor. Across nine feature-extractor/model combinations, the combination of a pre-trained ResNet50 feature extractor with a support vector machine gave the lowest prediction error on the paper's dataset, with an MSE of 0.21 and an MAE of 0.32. The paper interprets this as evidence that deep-learning features are more effective than hand-crafted green-area or color-texture features for capturing the complexity of grapevine canopies. If this holds, it matters because LAI is a standard indicator of crop health and growth, and the current manual or destructive measurement methods are slow, costly, and small-scale.","feed_headline":"ResNet plus SVM reads grapevine leaf area from drone photos","feed_subtitle":"Best of nine tested pipelines, with an MSE of 0.21, replacing slow manual LAI sampling with a fast non-destructive scan.","key_machinery":"The load-bearing component is the pre-trained ResNet50 convolutional network used as an image-to-feature extractor: for every cropped plant image it produces a feature vector that is then fed to the regressors. ResNet50 is a deep residual network pretrained on large-scale image data, and here it is not fine-tuned; it is used only to encode the image. The support vector regression with a radial basis function kernel is the second half of the winning pipeline, and the paper credits its kernel mapping for handling the complex, non-linear relations in the features. The same feature vectors are also pushed through linear regression and random forest, making the feature extractor the variable that is being compared across pipelines.","core_discovery":"The central claim is that among three feature extraction methods (a green-area edge-detection pipeline, a color-texture vocabulary built from color histograms, Hu moments, and Haralick textures, and a pre-trained ResNet50 deep network) paired with three regressors (linear regression, support vector regression with an RBF kernel, and a 100-tree random forest), the ResNet50-plus-SVM pipeline outperforms the others on the grapevine LAI dataset, reporting the lowest MSE (0.21) and MAE (0.32). The paper also observes that simpler green-area features remain competitive, especially with a random forest, and that linear regression is the least stable across feature sets. The reported superiority of the deep-learning pipeline is attributed to the ResNet feature extractor's ability to represent leaf texture, shape, and health in a way that the SVM can map to LAI.","pith_inferences":["A likely reason the green-area features stay competitive is that LAI is strongly correlated with canopy cover, so a domain-specific color-segmentation feature captures most of the signal; that interpretation goes beyond what the paper states.","Because each of the 1,469 plant images was cropped from one of only 498 drone frames, a rigorous evaluation should group crops by source image; this paper does not report such grouping, so its error numbers may overstate accuracy in a way the authors did not intend.","The same pretrained-features-plus-SVR recipe could plausibly transfer to other row crops with similar canopy geometry, but that is an untested extension rather than a claim the paper makes.","A direct practical check would be to see whether ResNet-plus-SVM retains its advantage when models are trained on one vineyard block or date and tested on another held-out block or date."],"forward_implications":["Grapevine LAI in a vineyard would be measurable from drone RGB images without destructive sampling, at the scale of individual plants.","Deep-learning feature extraction plus support vector regression is the best-performing recipe tested, but it is not the only workable one: green-area features with random forest are nearly as accurate.","A fast, cost-effective LAI pipeline of this kind could support precision agriculture tasks such as growth tracking and health assessment in small or resource-limited farms.","The method's usefulness depends on the reported accuracy transferring to new fields and seasons, since the comparison is made on one dataset of 1,469 cropped plant images."],"supporting_citations":[{"why":"Supplies the rationale for using pre-trained deep-learning features to improve machine-learning models in agriculture.","marker":"[5]"},{"why":"Grounds the overall choice of machine-learning regression models for agricultural prediction tasks.","marker":"[6]"},{"why":"Inspires the color-index and texture vocabulary feature extraction and its use for crop LAI estimation.","marker":"[7]"},{"why":"Establishes prior work using deep learning for crop LAI inversion that this study extends to grapevines.","marker":"[8]"},{"why":"Provides the canopy-analysis method that produced the measured LAI values used as training labels.","marker":"[11]"},{"why":"Supplies the Canny edge-detection technique used in the green-area feature extraction pipeline.","marker":"[12]"}],"fun_headline_variants":["ResNet-SVM on drone photos reads grapevine leaf area fast","Drone ResNet-SVM speeds up grapevine LAI measurement","Grapevine LAI from drone imagery: ResNet-SVM beats traditional","Deep learning+SVM estimates grapevine leaf area non-destructively","UAV imagery + ResNet-SVM gives quick grapevine leaf area"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported error values were measured on data the models had not already been fitted to; the paper never documents a train/test split or cross-validation scheme, so the best-pipeline result could be in-sample fit rather than predictive accuracy.","fun_headline_variants_meta":{"raw":{"variants":["ResNet-SVM on drone photos reads grapevine leaf area fast","Drone ResNet-SVM speeds up grapevine LAI measurement","Grapevine LAI from drone imagery: ResNet-SVM beats traditional","Deep learning+SVM estimates grapevine leaf area non-destructively","UAV imagery + ResNet-SVM gives quick grapevine leaf area"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001057,"raw_usage":{"total_tokens":4380,"prompt_tokens":837,"completion_tokens":3543,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":3449}},"tokens_in":453,"tokens_out":3543,"duration_ms":24715,"temperature":1.0,"reasoning_tokens":3449,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:43:00.390070+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run all nine feature-extractor/model combinations with a documented group-wise split that keeps every crop extracted from the same drone image inside the same training or test fold, and compare ResNet-plus-SVM against green-area-plus-random-forest; if the error gap closes or reverses, the claimed superiority of the deep-learning pipeline is an artifact of information leakage rather than genuine predictive quality.","supporting_citations":[{"cited_title":"A review of deep learning techniques used in agriculture","cited_arxiv_id":null,"evidence_quote":"Supplies the rationale for using pre-trained deep-learning features to improve machine-learning models in agriculture."},{"cited_title":"Machine learning in agriculture: A review","cited_arxiv_id":null,"evidence_quote":"Grounds the overall choice of machine-learning regression models for agricultural prediction tasks."},{"cited_title":"Combining color indices and textures of uav-based digital imagery for rice lai estimation","cited_arxiv_id":null,"evidence_quote":"Inspires the color-index and texture vocabulary feature extraction and its use for crop LAI estimation."},{"cited_title":"Esti- mating lai from winter wheat using uav data and cnns","cited_arxiv_id":null,"evidence_quote":"Establishes prior work using deep learning for crop LAI inversion that this study extends to grapevines."},{"cited_title":"The role of lai and leaf chlorophyll on ndvi estimated by uav in grapevine canopies","cited_arxiv_id":null,"evidence_quote":"Provides the canopy-analysis method that produced the measured LAI values used as training labels."},{"cited_title":"Extraction of yardang characteristics using object-based image analysis and canny edge detection methods","cited_arxiv_id":null,"evidence_quote":"Supplies the Canny edge-detection technique used in the green-area feature extraction pipeline."}],"review_version":1}