Pith. sign in

REVIEW 3 major objections 4 minor 12 references

Automating grapevine LAI features estimation with UAV imagery and machine learning

T0 review · 3 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A ResNet-plus-SVM pipeline estimates grapevine leaf area from drone RGB images with the lowest error of the nine combinations tested.

desk verdict Modest, useful grapevine LAI dataset and feature comparison, but the missing train/test protocol makes the headline numbers not yet trustworthy. read the letter →

arxiv 2411.17897 v1 pith:SAOFNOUN submitted 2024-11-26 cs.CV cs.AIcs.ETcs.LG

Muhammad Waseem Akram , Marco Vannucci , Giorgio Buttazzo , Valentina Colla , Stefano Roccella , Andrea Vannini , Giovanni Caruso , Simone Nesi
show 2 more authors
Alessandra Francini Luca Sebastiani
This is my paper · ORCID
classification cs.CVcs.AIcs.ETcs.LG
keywords LeafAreaIndexgrapevineUAVimagerymachinelearningregressiondeepfeatureextractionResNet50supportvectorprecisionagriculture
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

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.

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 (3)
  1. [Section III, Table I] 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.
  2. [Section II-A] 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.
  3. [Section II-C] 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.
minor comments (4)
  1. [Throughout] 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').
  2. [Section II-C] 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.
  3. [Section II-A] 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.
  4. [Figure 3] 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.

Circularity Check

2 steps flagged · score 6.0 of 10

The central ResNet+SVM accuracy claim is presented as prediction, but no held-out evaluation is described and feature thresholds are tuned to 'best results', so the reported errors reduce to in-sample fit.

  1. fitted input called prediction [Section II-C, Green Area Features Extraction using an Edge Detection]
    "Threshold values of 50 and 150 are used for this method. ... The reported values for all images are consistent and correspond to the best results."

    The threshold and preprocessing values in this pipeline are selected because they produce the best results on the same dataset, i.e., they are tuned with knowledge of the LAI target. When Table I later reports MSE/MAE for this feature set on that same dataset, the performance is partly a consequence of target-aware tuning rather than an independent test of the feature extractor.

  2. fitted input called prediction [Section III, Results and Discussion, Table I (with Section II-G Model Implementation)]
    "We employed the different parameters specified in the Scikit-Learn library to train the models. ... Performances are quantified in terms of Mean Squared Error (MSE), Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE). ... The achieved results show that the combination of SVM with ResNet outperforms other models, as it shows the lowest MSE (0.21) and MAE (0.32) values."

    No train/test split, cross-validation, or held-out subset is described anywhere; the only evaluation statement is in the Results section and the only implementation statement says parameters were used 'to train the models.' Thus the reported MSE/MAE are computed on the same data used to fit the models, so the 'prediction' errors in Table I reduce to in-sample residuals. The claim that ResNet+SVM is superior is therefore not an out-of-sample predictive result; it is a fit-quality comparison unless a held-out protocol is supplied.

full rationale

The paper is not circular in the equation-level sense: LAI labels come from an external SunScan instrument, features are obtained by standard pipelines (Equations 1-10), and the machine-learning models are conventional regressors. There is no self-citation chain or imported uniqueness theorem that forces the conclusion. The circularity concern is confined to the evaluation protocol. Section II-C explicitly tunes the Green Area thresholds to produce 'the best results' on this dataset, and Sections II-G and III provide no train/test split or cross-validation; Table I's MSE/MAE are therefore indistinguishable from in-sample residuals. Because the headline result ('SVM with ResNet ... lowest MSE (0.21) and MAE (0.32)') is claimed as predictive superiority but rests on metrics that appear to be computed on the fitted data, the central 'prediction' reduces, as reported, to goodness-of-fit. Crop-level leakage from 1,469 crops derived from 498 images is an additional risk that would only inflate such in-sample-style optimism. Hence the score is 6: partial circularity in the central empirical claim, not a definitional collapse or a self-citation chain.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new entities, forces, or conserved quantities. The central claim rests on empirical assumptions: accurate ground-truth LAI from SunScan, correct annotation correspondence, RGB content as a sufficient LAI signal, and transferability of ImageNet features. The free parameters are feature extraction thresholds, unreported SVM hyperparameters, RF tree count, and an undisclosed train/test split. The absence of a described evaluation protocol is the largest hidden parameter.

free parameters (5)
  • Green-area pixel threshold T=50 = 50 (all pixel values above 50 set to 255)
    This threshold is chosen to enhance vegetation detection and is described in Section II-C as corresponding to 'the best results' on the dataset, so it appears to be tuned on the same data used for evaluation.
  • Canny edge detection thresholds = 50 and 150
    The Canny thresholds are set in Section II-C and are part of the green-area feature pipeline that the authors say was selected to give best results; no independent validation is reported.
  • SVM RBF hyperparameters = scikit-learn defaults, not reported
    Section II-G says SVM uses the RBF kernel with parameters from scikit-learn, but specific C, epsilon, and gamma values are not reported. These choices affect the MSE values in Table I.
  • Random forest number of trees = 100
    Section II-F states the RF model uses 100 trees, chosen for 'optimizing performance while maintaining computational efficiency'; no sensitivity analysis is provided.
  • Train/test split and random seed = unreported
    The paper never specifies how the 1,469 images were divided for training and testing. This is a critical choice that determines whether the reported errors are predictive or in-sample.
assumptions (4)
  • domain assumption SunScan SS1-R3-BF3 canopy analyzer measurements are an accurate ground truth for individual plant LAI.
    Section II-A uses the SunScan system as reference without onsite calibration against destructive LAI measurements in this experiment. Any bias in SunScan readings propagates directly into model targets and errors.
  • domain assumption The manually drawn bounding boxes in aerial images correctly match the same plants that were measured by SunScan, with no spatial or temporal misregistration.
    Section II-A describes using commercial annotation software to associate each LAI value to a plant image. If a bounding box does not correspond to the measured plant, the image-label pair is wrong and the supervised learning is corrupted.
  • domain assumption Top-view RGB imagery of individual grapevine plants contains enough signal to estimate LAI.
    The entire feature extraction and regression pipeline assumes LAI is recoverable from RGB canopy appearance. This is plausible but not independently verified against occlusion, row structure, or shadow effects in this dataset.
  • domain assumption ImageNet-pretrained ResNet50 features transfer usefully to grapevine canopy images.
    Section II-E extracts features from a pretrained ResNet50 without fine-tuning or evidence that ImageNet representations are relevant to leaf density. The success of the approach depends on this transferability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Automating grapevine LAI features estimation with UAV imagery and machine learning." pith.science (2026). https://pith.science/paper/SAOFNOUN

@misc{pith2026241117897,
  author       = {Pith},
  title        = {Pith review of: Automating grapevine LAI features estimation with UAV imagery and machine learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SAOFNOUN}},
  note         = {Machine review of arXiv:2411.17897}
}
read the original abstract

The leaf area index determines crop health and growth. Traditional methods for calculating it are time-consuming, destructive, costly, and limited to a scale. In this study, we automate the index estimation method using drone image data of grapevine plants and a machine learning model. Traditional feature extraction and deep learning methods are used to obtain helpful information from the data and enhance the performance of the different machine learning models employed for the leaf area index prediction. The results showed that deep learning based feature extraction is more effective than traditional methods. The new approach is a significant improvement over old methods, offering a faster, non-destructive, and cost-effective leaf area index calculation, which enhances precision agriculture practices.

Figures

Figures reproduced from arXiv: 2411.17897 by the authors.

Figure 1
Figure 1. Examples a and b are annotated images with bounding [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Examples a,b,c and d are images showing individual plants [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the Leaf Area Estimation process using drone images and Machine Learning approaches [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

  1. [1]

    An overview of global leaf area index (lai): Methods, products, validation, and applications

    Hongliang Fang, Frederic Baret, Stephen Plummer, and Gabriela Schaepman-Strub. An overview of global leaf area index (lai): Methods, products, validation, and applications. Reviews of Geophysics, 57(3):739– 799, 2019

    work page 2019

  2. [2]

    Applications of remote sensing in precision agriculture: A review

    Rajendra P Sishodia, Ram L Ray, and Sudhir K Singh. Applications of remote sensing in precision agriculture: A review. Remote sensing , 12(19):3136, 2020

    work page 2020

  3. [3]

    Dynamic mapping of rice growth parameters using hj-1 ccd time series data

    Jing Wang, Jingfeng Huang, Ping Gao, Chuanwen Wei, and Lamin R Mansaray. Dynamic mapping of rice growth parameters using hj-1 ccd time series data. Remote Sensing, 8(11):931, 2016

    work page 2016

  4. [4]

    Integration of a crop growth model and deep learning methods to improve satellite-based yield estimation of winter wheat in henan province, china

    Yi Xie and Jianxi Huang. Integration of a crop growth model and deep learning methods to improve satellite-based yield estimation of winter wheat in henan province, china. Remote Sensing, 13(21):4372, 2021

    work page 2021

  5. [5]

    A review of deep learning techniques used in agriculture

    Ishana Attri, Lalit Kumar Awasthi, Teek Parval Sharma, and Priyanka Rathee. A review of deep learning techniques used in agriculture. Ecological Informatics, page 102217, 2023

    work page 2023

  6. [6]

    Machine learning in agriculture: A review

    Konstantinos G Liakos, Patrizia Busato, Dimitrios Moshou, Simon Pear- son, and Dionysis Bochtis. Machine learning in agriculture: A review. Sensors, 18(8):2674, 2018

    work page 2018

  7. [7]

    Combining color indices and textures of uav-based digital imagery for rice lai estimation

    Songyang Li, Fei Yuan, Syed Tahir Ata-UI-Karim, Hengbiao Zheng, Tao Cheng, Xiaojun Liu, Yongchao Tian, Yan Zhu, Weixing Cao, and Qiang Cao. Combining color indices and textures of uav-based digital imagery for rice lai estimation. Remote Sensing, 11(15):1763, 2019

    work page 2019

  8. [8]

    Esti- mating lai from winter wheat using uav data and cnns

    Lucas Wittstruck, Thomas Jarmer, Dieter Trautz, and Bj ¨orn Waske. Esti- mating lai from winter wheat using uav data and cnns. IEEE Geoscience and Remote Sensing Letters , 19:1–5, 2022

    work page 2022

Show all 12 references
  1. [9]

    Modeling maize above-ground biomass based on machine learning approaches using uav remote-sensing data

    Liang Han, Guijun Yang, Huayang Dai, Bo Xu, Hao Yang, Haikuan Feng, Zhenhai Li, and Xiaodong Yang. Modeling maize above-ground biomass based on machine learning approaches using uav remote-sensing data. Plant methods, 15:1–19, 2019

    2019

  2. [10]

    Environment 4.0: How digitalization and machine learning can improve the environmental footprint of the steel production processes

    Valentina Colla, Costanzo Pietrosanti, Enrico Malfa, and Klaus Peters. Environment 4.0: How digitalization and machine learning can improve the environmental footprint of the steel production processes. Materiaux et Techniques, 108(5-6), 2020

    2020

  3. [11]

    The role of lai and leaf chlorophyll on ndvi estimated by uav in grapevine canopies

    Giovanni Caruso, Giacomo Palai, Letizia Tozzini, Claudio D’Onofrio, and Riccardo Gucci. The role of lai and leaf chlorophyll on ndvi estimated by uav in grapevine canopies. Scientia Horticulturae, 322:112398, 2023

    2023

  4. [12]

    Extraction of yardang characteristics using object-based image analysis and canny edge detection methods

    Weitao Yuan, Wangle Zhang, Zhongping Lai, and Jingxiong Zhang. Extraction of yardang characteristics using object-based image analysis and canny edge detection methods. Remote Sensing, 12(4):726, 2020

    2020

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.