Pith. sign in

REVIEW 4 major objections 4 minor 43 references

MRD-LiNet: A Novel Lightweight Hybrid CNN with Gradient-Guided Unlearning for Improved Drought Stress Identification

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read The paper claims a 0.231M-parameter hybrid CNN detects potato drought stress at 90.0% accuracy, and that gradient-guided removal of 5% low-influence training images improves accuracy and reduces missed stressed plants.

desk verdict A credible lightweight-CNN paper whose efficiency claim is real but whose accuracy comparison and unlearning gain are both under-supported. read the letter →

arxiv 2509.06367 v1 pith:6EJXDW5C submitted 2025-09-08 cs.CV cs.AIcs.LG

classification cs.CVcs.AIcs.LG
keywords droughtstressidentificationlightweightCNNmachineunlearninggradientinfluencescoreprecisionagriculturepotatocropaerialimageryparameterefficiency
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 claims that accurate drought-stress identification does not need large, parameter-heavy models. A hybrid CNN combining bottleneck residual blocks, a dense feature-reuse block, and a transition layer reaches 90.0% test accuracy on an aerial potato dataset with only 0.231 million trainable parameters, within 1.6 points of a 14-million-parameter vision transformer baseline. It further claims that scoring each training image by the L2 norm of gradients of the model's prediction with respect to the input, then removing the lowest-scoring 5% and retraining, improves accuracy from 88.6% to 90.0% and lowers false negatives for stressed plants. If true, the practical payoff is real-time drought monitoring on drones and edge devices.

What carries the argument

Two mechanisms carry the argument. The architecture combines bottleneck residual blocks (1x1 expansion, 3x3 depthwise convolution, 1x1 projection, and an optional skip connection), a four-layer dense block that concatenates every layer's output with all preceding feature maps, and a transition layer that compresses channels with a 1x1 convolution and 2x2 average pooling; global average pooling feeds a 128-unit dense layer and a sigmoid output. The unlearning mechanism is a gradient-norm influence score: after training, each image's score is the L2 norm of the flattened gradient of the model's prediction with respect to that input image; the lowest-scoring 5% of the training set is discarded

What would settle it

Repeat the Aug+MU protocol with a random 5% of training images removed, replicated over several seeds. If random removal gives the same ~90.0% accuracy and similar stressed-class false-negative counts, the gradient-guided selection is not responsible for the gain. A sharper check: remove 2.5%, 7.5%, and 10% and see whether the gain tracks the influence ranking rather than the removal fraction.

Watch

Extended reading notes

Core claim

The central claim is twofold. First, a custom lightweight hybrid CNN—bottleneck residual blocks in the style of MobileNetV2, a four-layer dense block in the style of DenseNet, a channel-compressing transition layer, and global average pooling—can classify healthy versus drought-stressed potato patches at 90.0% accuracy with only 0.231M trainable parameters. This sits close to the 91.6% of a 14M-parameter transfer-learned vision transformer and above the 88.7% MobileNet pipeline, with a 15-60x parameter reduction. Second, the paper defines an influence score for each image as the L2 norm of the concatenated gradients of the model's prediction with respect to the input image, removes the 5% of

Load-bearing premise

The load-bearing premise is that deleting the 5% of training images with the lowest gradient-norm influence scores is what improves generalization; the paper does not compare with random deletion, so the measured gain could come from sample reduction or chance rather than from the influence ranking.

Editorial extensions

If this is right

  • Drought-stress screening can run on UAVs and edge devices: 0.231M parameters means a small memory footprint and fast inference, with accuracy within 1.6 points of the best heavy baseline.
  • Parameter efficiency does not have to come at the cost of unbalanced errors; the reported stressed-class recall and F1 improve with augmentation plus unlearning, reducing missed stressed plants.
  • Gradient-guided unlearning can be used as a data-cleaning step in agricultural datasets, letting a deployed model adapt by forgetting low-influence or noisy samples.
  • The architecture recipe—combining bottleneck residuals, a dense block, and a transition layer—is a viable template for resource-limited image classification beyond drought detection.

Reading between the lines

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

  • Extension: the unlearning gain is not yet isolated from plain data reduction. The paper lacks a random-removal baseline, so the 1.4-point accuracy gain could be produced by dropping any 5% of the training set.
  • Extension: the influence score measures sensitivity of the prediction to input pixels, not the sample's contribution to the loss; what the paper calls 'forgetting' may amount to pruning outliers or easy redundant images rather than privacy-style unlearning.
  • Extension: because the test set comes from a single potato field and season, the 90.0% figure should be checked on other fields, crops, growth stages, and sensor conditions before treating it as a general drought-detection result.
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

4 major / 4 minor

Summary. The paper proposes MRD-LiNet, a lightweight hybrid CNN combining MobileNetV2-style bottleneck residual blocks, a DenseNet-like dense block, and a transition layer, with 0.231M trainable parameters. It also introduces a machine-unlearning step that computes per-sample influence scores from gradients, removes the 5% least influential training samples, and retrains the model. The method is evaluated on a potato drought-stress aerial dataset with 60 test images and 1,135 annotated windows. The authors report accuracies of 88.1% (no augmentation), 88.6% (augmentation), and 90.0% (augmentation plus unlearning), and compare these against three prior pipelines (MobileNet, DenseNet121, ViT-TL), claiming a 15–60-fold parameter reduction and competitive accuracy.

Significance. If the empirical claims hold, the parameter-efficiency result (0.231M parameters at 90.0% accuracy) is practically useful for UAV/edge deployment, and the use of gradient-guided unlearning to improve generalization is an interesting direction. The paper does not, however, ship machine-checked proofs, reproducible code, or a random-removal control; the main empirical contributions are supported only by single-run comparisons. The comparison against prior work is also not demonstrated to be protocol-matched. The central ideas are therefore worthy of consideration, but the evidence as presented is insufficient for acceptance.

major comments (4)
  1. [§4.1, Scenario (iii)] The only evidence for the unlearning benefit is the 1.4-point gap between Aug (88.6%) and Aug+MU (90.0%) on 1,135 test windows, i.e., roughly 16 correctly classified samples. There is no random-removal control, no repeated runs, and no error bars. Without a random-5%-removal baseline and multiple seeds, the improvement cannot be attributed to gradient-guided selection rather than data reduction or chance. Please add the control and report mean±std.
  2. [§3.1 vs §3.1.1] The influence score is defined inconsistently. The displayed formula defines it as the norm of the gradient of the loss with respect to the model prediction, while §3.1.1 and Algorithm 2 compute the L2 norm of the gradient of the prediction with respect to the input image, flattened and concatenated across layers. These are different quantities. The manuscript must state exactly which gradient was computed and how the score was aggregated; otherwise the unlearning mechanism is not reproducible.
  3. [Table 3 and §4.2] The accuracy comparison is not shown to be protocol-matched. Baselines [13] and [26] are previous papers by the same group, and the text does not establish that they used the same 60-image/1,135-window test split, same preprocessing, same augmentation, or same evaluation script. As reported, the 0.7–1.6 point gaps to DenseNet121/ViT-TL are not interpretable. Additionally, §4.2 claims the proposed model has the lowest stressed-class false-negative rate (13.5%), but Table 3 shows ViT-TL has stressed recall 0.901, i.e., 9.9% false negatives. This claim is factually wrong as stated.
  4. [Algorithm 1 vs §3.0.1] The architecture specification is internally inconsistent: §3.0.1 states the initial convolution uses 32 filters, while Algorithm 1 sets Conv2D(36,(3,3)). The exact filter count affects the parameter total, which is a headline result. Please reconcile the text and algorithm, and state the precise parameter-count calculation. A public code release would also resolve this ambiguity.
minor comments (4)
  1. [§3.0.6 / §4] The loss function is called 'binary crossentropy' in §3.0.6 and 'categorical cross-entropy' in §4. For a single-sigmoid binary output, binary crossentropy is the correct name; please fix.
  2. [Table 2] The table formatting merges healthy-class precision, recall, and F1 into a single cell (e.g., '0.760.96 * 0.85'), making it hard to read. Please separate the columns and explain the asterisk.
  3. [§3.1] Typo: 'calulated' should be 'calculated'.
  4. [§4.1] The removal fraction (5%) is presented as fixed without sensitivity analysis. A short sweep (e.g., 1%, 5%, 10%) would strengthen the claim that this choice is not arbitrary.

Circularity Check

1 steps flagged · score 4.0 of 10

Competitive-accuracy claim rests on self-cited baselines without shared evaluation protocol; absolute efficiency results are independent.

  1. self citation load bearing [Section 4.2, Table 3 and surrounding text]
    "Prior pipelines such as MobileNet [13] and DenseNet121 [13] achieved respectable accuracies of 88.7% and 90.7%, respectively, while the ViT-TL model [26] reached the highest reported accuracy of 91.6%. ... In contrast, the proposed framework attains a competitive accuracy of 90.0% ..."

    The 'competitive accuracy' claim is established by comparing to accuracies from [13] and [26], which are prior works by the current authors. The paper does not state that these baselines were retrained on the same test set (60 images, 1,135 windows) or with the same preprocessing, so the comparison reduces to trusting the authors' own prior results. The proposed model's absolute accuracy and parameter count are from the present experiment, so the efficiency claim remains independent; only the relative 'competitive' assertion is self-citation-dependent.

full rationale

The paper's core contribution is an empirical hybrid CNN with a gradient-norm-based data-pruning step; no equation reduces to its own input. However, the headline claim of 'competitive accuracy' is supported only by Table 3, which lists accuracies from [13] and [26]—both authored by the current paper's authors—without evidence of a shared test split or protocol. This is a load-bearing self-citation for the relative performance claim. The absolute results (90.0% accuracy, 0.231M parameters) and the internal Aug vs Aug+MU comparison are self-contained and do not exhibit circularity. The lack of a random-removal baseline for the unlearning step is a scientific weakness, but not a circularity. Overall score 4: some self-citation, central efficiency claim has independent content.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on hand-chosen architecture hyperparameters, a hand-picked unlearning removal fraction, and untested assumptions about annotation quality, influence-score validity, and baseline comparability. No new physical or conceptual entities are introduced.

free parameters (3)
  • Unlearning removal fraction = 5% of training samples
    Chosen by hand in Section 4.1; no sensitivity analysis (e.g., 2%, 10%) is reported, so the effect of this choice on the results is unknown.
  • Initial convolutional filters = 32 (text) vs 36 (Algorithm 1)
    The architecture description is inconsistent; in either case the filter count is a design choice, not derived from data or theory.
  • Dense block growth rate and number of layers = growth rate 32, 4 layers
    Specified in Algorithm 1; chosen without ablation, and these values directly determine the 0.231M parameter count and the model's capacity.
assumptions (4)
  • domain assumption Manual annotations of healthy and stressed potato regions are correct and representative.
    Section 2.1: annotations were made manually using LabelImg; no inter-annotator agreement or annotation error analysis is reported, so the ground truth is taken on faith.
  • ad hoc to paper The L2 norm of gradients with respect to input (or loss) is a valid measure of training sample influence for the purpose of pruning.
    Section 3.1: the influence score is defined as a gradient norm, but no theoretical justification or empirical comparison to other influence heuristics is provided, and the score is not validated against a random-removal baseline.
  • domain assumption The 60-image test set with 1,135 windows yields a reliable accuracy estimate.
    The test set is a single split with no confidence intervals, no repeated train/test resampling, and no statistical significance testing.
  • domain assumption Results in Table 3 from prior papers are directly comparable to the proposed model's results.
    Table 3 quotes numbers from [13] and [26] (both involving the current authors) without re-running those models on the same train/test split, so the '15-fold reduction' and 'competitive accuracy' claims assume split comparability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MRD-LiNet: A Novel Lightweight Hybrid CNN with Gradient-Guided Unlearning for Improved Drought Stress Identification." pith.science (2026). https://pith.science/paper/6EJXDW5C

@misc{pith2026250906367,
  author       = {Pith},
  title        = {Pith review of: MRD-LiNet: A Novel Lightweight Hybrid CNN with Gradient-Guided Unlearning for Improved Drought Stress Identification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6EJXDW5C}},
  note         = {Machine review of arXiv:2509.06367}
}
read the original abstract

Drought stress is a major threat to global crop productivity, making its early and precise detection essential for sustainable agricultural management. Traditional approaches, though useful, are often time-consuming and labor-intensive, which has motivated the adoption of deep learning methods. In recent years, Convolutional Neural Network (CNN) and Vision Transformer architectures have been widely explored for drought stress identification; however, these models generally rely on a large number of trainable parameters, restricting their use in resource-limited and real-time agricultural settings. To address this challenge, we propose a novel lightweight hybrid CNN framework inspired by ResNet, DenseNet, and MobileNet architectures. The framework achieves a remarkable 15-fold reduction in trainable parameters compared to conventional CNN and Vision Transformer models, while maintaining competitive accuracy. In addition, we introduce a machine unlearning mechanism based on a gradient norm-based influence function, which enables targeted removal of specific training data influence, thereby improving model adaptability. The method was evaluated on an aerial image dataset of potato fields with expert-annotated healthy and drought-stressed regions. Experimental results show that our framework achieves high accuracy while substantially lowering computational costs. These findings highlight its potential as a practical, scalable, and adaptive solution for drought stress monitoring in precision agriculture, particularly under resource-constrained conditions.

Figures

Figures reproduced from arXiv: 2509.06367 by the authors.

Figure 1
Figure 1. Field images showing a) Sample RGB image and b) Healthy and Stressed Labels [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Schematic diagram of the lightweight network architecture The proposed framework begins with an initial con￾volutional layer for low-level feature extraction, followed by four residual blocks that enable deeper feature learning through skip connections. A dense block is then employed to encourage feature reuse and efficient representation learning, after which a transition layer reduces dimensionality and controls m… view at source ↗
Figure 3
Figure 3. Machine Unlearning Framework Algorithm 2: Calculate Influence Scores Input: Trained model 𝑀, DataFrame 𝐷 with image filenames and labels Output: Array of influence scores 𝑆 1 Initialize empty list 𝑆 ← [ ]; 2 foreach row 𝑟 in 𝐷 do 3 𝑝 ← filename from 𝑟; 4 𝑦 ← label from 𝑟 converted to float32; 5 𝑋 ← PreprocessImageForCustomCNN(𝑝); 6 𝐺 ← ComputeGradients(𝑀, 𝑋, ExpandDims(𝑦)); 7 Initialize empty list 𝐹; 8 foreach gradi… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Distribution of influence scores for training samples (50 bins). (a) No Augmentation (b) With Augmentation (c) With Augmentation + MU (5%) [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Learning curves (accuracy and loss) for the CNN under three scenarios: (a) without augmentation, (b) with augmentation, and (c) with augmentation + machine unlearning (5% data removal). Each subfigure shows Accuracy (top) and Loss (bottom). [26] reached the highest rep…
Figure 6
Figure 6. Figure 6: Confusion matrices for the three scenario [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 36 canonical work pages

  1. [13]

    Explainable light-weight deep learning pipeline for improved drought stress identification

    Aswini Kumar Patra and Lingaraj Sahoo. Explainable light-weight deep learning pipeline for improved drought stress identification. Frontiers in Plant Science, 15:1476130, 2024

  2. [26]

    An explainable vision transformer with transfer learning based efficient drought stress identification.Plant Molecular Biology, 115(4):98, 2025

    Aswini Kumar Patra, Ankit Varshney, and Lingaraj Sahoo. An explainable vision transformer with transfer learning based efficient drought stress identification.Plant Molecular Biology, 115(4):98, 2025

  3. [1]

    Jose Luis Araus and Jill E. Cairns. Field high-throughput phenotyp- ing:thenewcropbreedingfrontier.TrendsinPlantScience,19(1):52– 61, January 2014

  4. [2]

    Next-generation breeding strategies for climate- ready crops.FRONTIERS IN PLANT SCIENCE, 12, JUL 21 2021

    Ali Razzaq, Parwinder Kaur, Naheed Akhter, Shabir Hussain Wani, and Fozia Saleem. Next-generation breeding strategies for climate- ready crops.FRONTIERS IN PLANT SCIENCE, 12, JUL 21 2021

  5. [3]

    Farooq, A

    M. Farooq, A. Wahid, N. Kobayashi, D. Fujita, and S. M. A. Basra. Plant Drought Stress: Effects, Mechanisms and Management. In Eric Lichtfouse, Mireille Navarrete, Philippe Debaeke, Souchere Veronique, and Caroline Alberola, editors,Sustainable Agriculture, pages 153–188. Springer Netherlands, Dordrecht, 2009

  6. [4]

    Kamilaris and F

    A. Kamilaris and F. X. Prenafeta-Boldú. A review of the use of convolutional neural networks in agriculture.The Journal of Agricultural Science, 156(3):312–322, April 2018

  7. [5]

    Computers and Electronics in Agriculture, 140:461–468, 2017

    Shuo Zhuang, Ping Wang, Boran Jiang, Maosong Li, and Zhihong Gong.Earlydetectionofwaterstressinmaizebasedondigitalimages. Computers and Electronics in Agriculture, 140:461–468, 2017

  8. [6]

    IdentificationandClassificationofMaizeDroughtStressUsing Deep Convolutional Neural Network.Symmetry, 11(2):256, 2019

    Jiangyong An, Wanyi Li, Maosong Li, Sanrong Cui, and Huanran Yue. IdentificationandClassificationofMaizeDroughtStressUsing Deep Convolutional Neural Network.Symmetry, 11(2):256, 2019

Show all 43 references
  1. [7]

    Adduru U. G. Sankararao, P. Rajalakshmi, and Sunita Choudhary. Machine Learning-Based Ensemble Band Selection for Early Water Stress Identification in Groundnut Canopy Using UAV-Based Hy- perspectral Imaging.IEEE Geoscience and Remote Sensing Letters, 20:1–5, 2023

  2. [8]

    Locke, Steven Mirsky, and Edgar Lobaton

    Paula Ramos-Giraldo, Chris Reberg-Horton, Anna M. Locke, Steven Mirsky, and Edgar Lobaton. Drought Stress Detection Using Low- Cost Computer Vision Systems and Machine Learning Techniques. IT Professional, 22(3):27–29, 2020

  3. [9]

    Shiva Azimi, Rohan Wadhawan, and Tapan K. Gandhi. Intelligent monitoringofstressinducedbywaterdeficiencyinplantsusingdeep learning.IEEE Transactions on Instrumentation and Measurement, 70:1–13, 2021

  4. [10]

    Identify- ing crop water stress using deep learning models.Neural Computing and Applications, 33(10):5353–5367, 2021

    Narendra Singh Chandel, Subir Kumar Chakraborty, Yogesh Anand Rajwade,KumkumDubey,MukeshK.Tiwari,andDilipJat. Identify- ing crop water stress using deep learning models.Neural Computing and Applications, 33(10):5353–5367, 2021

  5. [11]

    Drought stress detection technique for wheat crop using machine learning.PeerJ Computer Science, 9:e1268, 2023

    Ankita Gupta, Lakhwinder Kaur, and Gurmeet Kaur. Drought stress detection technique for wheat crop using machine learning.PeerJ Computer Science, 9:e1268, 2023

  6. [12]

    Potatocropstressidentificationinaerialimages using deep learning-based object detection.Agronomy Journal, 113:3991–4002, 2021

    SujataButte,Vakanski,Aleksandar,KasiaDuellman,HaotianWang, andAminMirkouei. Potatocropstressidentificationinaerialimages using deep learning-based object detection.Agronomy Journal, 113:3991–4002, 2021

  7. [14]

    Hyperspectral machine-learning model for screening tea germplasm resources with drought tolerance.Frontiers in Plant Science, 13, 2022

    Sizhou Chen, Jiazhi Shen, Kai Fan, Wenjun Qian, Honglian Gu, Yuchen Li, Jie Zhang, Xiao Han, Yu Wang, and Zhaotang Ding. Hyperspectral machine-learning model for screening tea germplasm resources with drought tolerance.Frontiers in Plant Science, 13, 2022

  8. [15]

    Dao, Yuhong He, and Cameron Proctor

    Phuong D. Dao, Yuhong He, and Cameron Proctor. Plant drought impact detection using ultra-high spatial resolution hyperspectral imagesandmachinelearning.InternationalJournalofAppliedEarth Observation and Geoinformation, 102:102364, 2021

  9. [16]

    PoojaGoyal,RakeshSharda,MukeshSaini,andMukeshSiag.Adeep learningapproachforearlydetectionofdroughtstressinmaizeusing proximal scale digital images.Neural Computing and Applications, 36(4):1899–1913, 2024

  10. [17]

    An image is worth 16x16 words: Transformers for image recognition at scale.https://arxiv.org/abs/2010.11929v2,

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weis- senborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition ...

  11. [18]

    Yuille, and Yuyin Zhou

    Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L. Yuille, and Yuyin Zhou. TransUNet: Transformersmakestrongencodersformedicalimagesegmentation. http://arxiv.org/abs/2102.04306, 2021

  12. [19]

    Swin transformer: Hierarchical vision transformer using shifted windows.http://arxiv.org/abs/ 2103.14030, 2021

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows.http://arxiv.org/abs/ 2103.14030, 2021

  13. [20]

    Acustomisedvisiontransformer for accurate detection and classification of java plum leaf disease

    AuvickChandraBhowmik,Md.TaimurAhad,YousufRayhanEmon, FarukAhmed,BoSong,andYanLi. Acustomisedvisiontransformer for accurate detection and classification of java plum leaf disease. Smart Agricultural Technology, 8:100500, 2024

  14. [21]

    Effective plant disease diagnosis using vision transformer trained with leafy-generative adversarial network- generated images.Expert Systems with Applications, 254:124387, 2024

    Aadarsh Kumar Singh, Akhil Rao, Pratik Chattopadhyay, Rahul Maurya, and Lokesh Singh. Effective plant disease diagnosis using vision transformer trained with leafy-generative adversarial network- generated images.Expert Systems with Applications, 254:124387, 2024

  15. [22]

    A deep learning based approach for automated plant disease classification using vision transformer.Scientific Reports, 12(1):11554, 2022

    Yasamin Borhani, Javad Khoramdel, and Esmaeil Najafi. A deep learning based approach for automated plant disease classification using vision transformer.Scientific Reports, 12(1):11554, 2022

  16. [23]

    Vision transformer meets con- volutional neural network for plant disease classification.Ecological Informatics, 77:102245, 2023

    Poornima Singh Thakur, Shubhangi Chaturvedi, Pritee Khanna, Tanuja Sheorey, and Aparajita Ojha. Vision transformer meets con- volutional neural network for plant disease classification.Ecological Informatics, 77:102245, 2023

  17. [24]

    ViT- SmartAgri: Vision transformer and smartphone-based plant disease detection for smart agriculture.Agronomy, 14(2):327, 2024

    Utpal Barman, Parismita Sarma, Mirzanur Rahman, Vaskar Deka, Swati Lahkar, Vaishali Sharma, and Manob Jyoti Saikia. ViT- SmartAgri: Vision transformer and smartphone-based plant disease detection for smart agriculture.Agronomy, 14(2):327, 2024

  18. [25]

    Alanazi, and Jong Weon Lee

    Sana Parez, Naqqash Dilshad, Norah Saleh Alghamdi, Turki M. Alanazi, and Jong Weon Lee. Visual intelligence in precision agriculture: Exploring plant disease detection via efficient vision transformers.Sensors, 23(15):6949, 2023

  19. [27]

    TrIncNet:alightweightvisiontransformer networkforidentificationofplantdiseases.FrontiersinPlantScience, 14, 2023

    Pushkar Gole, Punam Bedi, Sudeep Marwaha, Md Ashraful Haque, andChandanKumarDeb. TrIncNet:alightweightvisiontransformer networkforidentificationofplantdiseases.FrontiersinPlantScience, 14, 2023

  20. [28]

    For- merLeaf: An efficient vision transformer for cassava leaf disease detection.Computers and Electronics in Agriculture, 204:107518, 2023

    Huy-Tan Thai, Kim-Hung Le, and Ngan Luu-Thuy Nguyen. For- merLeaf: An efficient vision transformer for cassava leaf disease detection.Computers and Electronics in Agriculture, 204:107518, 2023

  21. [29]

    A novel hierarchical framework for plant leaf disease detection using residual vision transformer

    Sasikala Vallabhajosyula, Venkatramaphanikumar Sistla, and Venkata Krishna Kishore Kolli. A novel hierarchical framework for plant leaf disease detection using residual vision transformer. Heliyon, 10(9), 2024

  22. [30]

    PMVT: a lightweight vision transformer for plant disease identificationonmobiledevices.FrontiersinPlantScience,14,2023

    Guoqiang Li, Yuchao Wang, Qing Zhao, Peiyan Yuan, and Baofang Chang. PMVT: a lightweight vision transformer for plant disease identificationonmobiledevices.FrontiersinPlantScience,14,2023

  23. [31]

    Resnet50

    BrettKoonce. Resnet50. InConvolutionalneuralnetworkswithswift for tensorflow: image recognition and dataset categorization, pages 63–72. Springer, 2021

  24. [32]

    Densenet: Implementing efficient convnet descriptor pyramids.arXiv preprint arXiv:1404.1869, 2014

    Forrest Iandola, Matt Moskewicz, Sergey Karayev, Ross Girshick, Trevor Darrell, and Kurt Keutzer. Densenet: Implementing efficient convnet descriptor pyramids.arXiv preprint arXiv:1404.1869, 2014

  25. [33]

    Mobilenets:Efficientconvolutionalneuralnetworksformobilevision applications.arXiv preprint arXiv:1704.04861, 2017

    Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, WeijunWang,TobiasWeyand,MarcoAndreetto,andHartwigAdam. Mobilenets:Efficientconvolutionalneuralnetworksformobilevision applications.arXiv preprint arXiv:1704.04861, 2017

  26. [34]

    Machineunlearning: Solutions and challenges.IEEE Transactions on Emerging Topics in Computational Intelligence, 8(3):2150–2168, 2024

    JieXu,ZihanWu,CongWang,andXiaohuaJia. Machineunlearning: Solutions and challenges.IEEE Transactions on Emerging Topics in Computational Intelligence, 8(3):2150–2168, 2024. 10

  27. [35]

    An overview of machine unlearning

    Chunxiao Li, Haipeng Jiang, Jiankang Chen, Yu Zhao, Shuxuan Fu, Fangming Jing, and Yu Guo. An overview of machine unlearning. High-Confidence Computing, 5(2):100254, 2025

  28. [36]

    Machine unlearning

    Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette- Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In2021 IEEE symposium on security and privacy (SP), pages 141–159. IEEE, 2021

  29. [37]

    Towards making systems forget with machine unlearning

    Yinzhi Cao and Junfeng Yang. Towards making systems forget with machine unlearning. In2015 IEEE symposium on security and privacy, pages 463–480. IEEE, 2015

  30. [38]

    Amnesiac ma- chine learning

    Laura Graves, Vineel Nagisetty, and Vijay Ganesh. Amnesiac ma- chine learning. InProceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 11516–11524, 2021

  31. [39]

    Eternal sunshine of the spotless net: Selective forgetting in deep networks

    Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9304–9312, 2020

  32. [40]

    Certified data removal from machine learning models

    Chuan Guo, Tom Goldstein, Awni Hannun, and Laurens Van Der Maaten. Certified data removal from machine learning models. arXiv preprint arXiv:1911.03030, 2019

  33. [41]

    Towards unbounded machine unlearning.Advances in neural information processing systems, 36:1957–1987, 2023

    Meghdad Kurmanji, Peter Triantafillou, Jamie Hayes, and Eleni Tri- antafillou. Towards unbounded machine unlearning.Advances in neural information processing systems, 36:1957–1987, 2023

  34. [42]

    A dataset of multispectral potato plants images, university of idaho.https://www.idahofallshighered.org/vakanski/Multispectral_ Images_Dataset.html, 2021

  35. [43]

    Labelimg.https://github.com/tzutalin/labelImg, 2019

    Tzutalin. Labelimg.https://github.com/tzutalin/labelImg, 2019. 11

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.