Pith. sign in

REVIEW 5 major objections 6 minor 62 references

Perturbation-Aware Diffusion-Guided Hybrid Segmentation for Robust and Annotation-Efficient Plant Stress Phenotyping

T0 review · 5 major / 6 minor · reviewed 2026-07-30 · grok-4.5

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

Pith's one-line read Matched backbone–diffusion refiner pairs plus perturbation-aware retraining improve plant stress masks under limited labels, compute, and domain shift.

desk verdict Useful applied recipe for plant stress masks under a 6 GB budget; the 71.83% headline and “beats PlantSegV3 reference” framing are not on a comparable protocol. read the letter →

arxiv 2607.23680 v1 pith:YVPKOE6C submitted 2026-07-26 cs.CV

Gurbhit Chaurakoti , Soumyashree Kar This is my paper
classification cs.CV
keywords DiffusionrefinementSemanticsegmentationAnnotation-efficientPerturbation-awaremodelingDomaininvariancePlantstressphenotypingBoundary-awareoptimization
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 argues that plant stress segmentation should be judged not only by clean in-domain scores but by how well masks hold up under appearance changes, scarce labels, and new fields. It builds hybrid models in which a standard segmenter first draws a coarse disease mask and a diffusion model then refines structure and boundaries. Across a staged pipeline on PlantSegV3—short architectural screening, boundary-focused training, low-data tests, failure-mode augmentation, and controlled transfer—it finds that only some backbone–refiner pairings help, that boundary-aware losses mainly sharpen contours, and that models stay near full-data quality even at about 10% labels. Perturbation ranking shows grayscale, fog, coarse dropout, and shadow hurt most, and retraining on those failures recovers robustness without wrecking clean performance. Adapted checkpoints also regain useful accuracy on external crop disease and crop–weed sets when a little target data is available, which the authors read as evidence that diffusion refinement plus boundary training supply transferable structural priors under realistic budgets.

What carries the argument

Diffusion-guided hybrid segmentation: a discriminative backbone (U-Net, DeepLabV3+, or SegFormer) produces a coarse mask that a diffusion refiner (DDPM, latent diffusion, or semantic-guided diffusion) corrects for boundary continuity and structural coherence, trained under a staged, resource-aware protocol rather than as a standalone segmenter.

What would settle it

Retrain the discarded pairings—especially SegFormer and latent-diffusion variants—for many more epochs with matched budgets and the same boundary and augmentation recipe; if any discarded hybrid then beats the selected models on refined mIoU and Boundary-F1 under clean, perturbed, low-data, and transfer tests, the matched-pairing claim from short screening fails.

Watch

Extended reading notes

Core claim

Carefully matched backbone–refiner pairings, combined with boundary-constrained optimization and perturbation-aware retraining, improve structural delineation and robustness for plant stress segmentation under limited compute, reduced annotation, and cross-domain shift. On PlantSegV3 the best selected hybrid reaches 71.83% refined mIoU and 26.10% refined Boundary-F1; the selected models remain stable near full-data levels at 10% labels; and controlled target adaptation recovers substantial performance on external agricultural datasets after direct transfer fails.

Load-bearing premise

A very short screening run on a small local GPU is assumed to fairly rank all nine hybrids, so discarding weaker pairs does not throw away setups that would win with longer or better-tuned training.

Editorial extensions

If this is right

  • Plant stress segmentation pipelines can treat diffusion as a structural corrector after a CNN backbone rather than as a full replacement segmenter.
  • Annotation budgets can be cut sharply (toward ~10% labels) when pretrained encoders and boundary-aware losses are kept, without large drops in mask quality on redundant field datasets.
  • Augmentation should be driven by measured failure modes (grayscale, fog, dropout, shadow) instead of generic photo jitter if robustness is the goal.
  • Source models trained this way carry structural priors that few-shot target fine-tuning can unlock on new disease and crop–weed datasets, even when zero-shot transfer collapses.
  • Useful model selection for field deployment can be done under tight GPU memory by short screens plus targeted deeper runs, not only by exhaustive search.

Reading between the lines

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

  • If short-budget ranking systematically underrates slower-converging transformers or latent refiners, the practical selection rule may need a two-pass screen: cheap filter then equalized longer fine-tunes of near-ties.
  • The grayscale failure and recovery pattern suggests color-invariant pretraining or explicit grayscale branches as a cheap next robustness lever beyond the paper’s augmentation curriculum.
  • Because CWFID adaptation must relearn label ontology as well as appearance, the same coarse-to-refine stack may transfer better to other lesion-style tasks than to full scene-semantic ag tasks without a label-space adapter.
  • Distilling the winning hybrids into lighter students, as the authors flag for future work, would test whether the structural prior survives compression for edge farm devices.
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

5 major / 6 minor

Summary. The manuscript presents a coarse-to-refined plant-disease segmentation pipeline in which U-Net, DeepLabV3+, or SegFormer backbones produce coarse masks that are refined by one of three diffusion modules (DDPM, latent diffusion, semantic-guided diffusion). A 3×3 screening study on PlantSegV3 (5-epoch runs on a 6 GB GPU) selects three hybrids, which are then subjected to boundary-constrained optimization (Dice + boundary + Lovász), low-data evaluation (100%→10% nested subsets), a perturbation sensitivity analysis whose ranking drives a two-phase augmentation policy, constrained hyperparameter screening, and few-shot cross-domain adaptation to NWRD and CWFID. The best model reports 71.83% refined mIoU and 26.10% Boundary-F1 on the PlantSegV3 validation split; low-data performance is stable at 10% labels (shown to depend on ImageNet pretraining, Table 10); augmentation retraining recovers much of the perturbation-induced drop; and adapted models partially recover on external datasets. The headline claims are (i) matched backbone–refiner pairings plus boundary-aware optimization improve delineation, and (ii) the pipeline is robust under limited compute, labels, and domain shift.

Significance. If the protocol issues are resolved, this is a useful, unusually thorough resource-aware study: the staged design connecting screening, boundary optimization, perturbation analysis, and failure-mode-derived augmentation is a sensible and reproducible-in-principle methodology; the pretraining ablation at 10% supervision (Table 10, +0.20 mIoU mean gain) is a genuinely informative control; and the perturbation ranking → augmentation policy loop (Tables 6–7) is a concrete, falsifiable contribution. The authors are commendably transparent about hardware constraints and mostly hedge their architectural conclusions to the short-training regime. However, all reported numbers are validation-set numbers under an unstated label space, no test-split result appears anywhere, and the one explicit comparison to the published PlantSegV3 benchmark is almost certainly not protocol-comparable. These are fixable within the manuscript's scope, but they currently gate the headline claim.

major comments (5)
  1. [§2.2, §3.9, Table 11, Eqs. (5)–(6)] The comparison underlying the headline number is not protocol-comparable as written. (a) The number of classes C in Eqs. (5)–(6) is never stated; the reported profile (mIoU ≈ 0.71, Dice ≈ 0.79, pixel accuracy ≈ 0.90 across all nine configs in Table 3) is arithmetically consistent with binary foreground/background segmentation, whereas the cited reference (ConvNeXt-L 46.24% mIoU, Wei et al. 2026) is computed on PlantSegV3's multi-class label space (115 plant–disease cases, §2.2). A ~25-point mIoU gain from adding a refinement module is not plausible under a fixed protocol. (b) The data are re-split: 7,916 train + 1,247 validation = 9,163 pairs, which exceeds the dataset's stated 7,774 images, and the original 70/10/20 split is not used — so it is literally not 'the same source-domain benchmark' as §3.9 claims. The authors must state C explicitly, explain the split arithmetic, and either e
  2. [§2.1, §3.1–3.6, Tables 3–8] Every reported number is a validation-set number, and checkpoint selection at each of the six sequential stages is by minimum validation loss on the same split (§2.5, Table 5 caption). Screening, boundary optimization, low-data, augmentation retraining, and hyperparameter screening thus accumulate adaptive selection pressure on a single split, and no result is ever reported on the official PlantSegV3 test set. The final selected checkpoints (Table 8 configurations) should be evaluated once on the held-out test split, and the abstract's 71.83%/26.10% figures should be test-set numbers or clearly relabeled as validation results.
  3. [§3.1, Table 3] The architectural conclusions rest on single 5-epoch screening runs with no seed variance reported anywhere in the paper. Several selection-relevant deltas are within plausible run-to-run noise (e.g., +0.0028 mIoU for config 1-1; negative Dice deltas in 1-1, 2-2, 3-1), and the discarded branches show signs of being disadvantaged by the budget rather than by architecture: SegFormer + LDM has the largest coarse→refined delta in the table (+0.0273) and transformer/latent-diffusion models are known to converge slowly. Either provide multi-seed statistics for the screening and a longer-budget control for at least one discarded branch, or scope the 'matched pairing' and 'latent diffusion is not consistently beneficial' claims explicitly to the 5-epoch regime (the text hedges in places but the abstract and conclusions do not).
  4. [§3.1–3.2, Tables 3–4] The value of the diffusion refiner relative to its cost is not established. Refined−coarse mIoU deltas are at most +0.0128 and sometimes negative, while no inference-time cost, number of denoising steps, noise schedule, or comparison to a cheap post-processing baseline (e.g., dense CRF or morphological refinement on the coarse mask) is reported. Since the paper's framing is explicitly compute-constrained ('limited compute' is a headline condition), evidence that the diffusion stage buys something a near-free post-processor does not is load-bearing for the central claim that diffusion refinement is the right mechanism.
  5. [§3.2, Table 4; Abstract] The Boundary-F1 story needs calibration. Stage-1 BF1 ≈ 0.06–0.08 implies near-random boundary localization before L_bco, and the ~3× jump to 0.22–0.25 from a loss change alone (no architectural change) warrants analysis: is the one-pixel tolerance band (§2.11) only satisfiable after explicit boundary supervision, and what does BF1 = 0.26 mean in absolute terms for delineation quality? Presenting 26.10% BF1 in the abstract as evidence of 'strong contour fidelity' overstates the absolute level; the claims should be tempered and the tolerance-band sensitivity (e.g., BF1 at 2–3 px) reported.
minor comments (6)
  1. [§3.6 vs Table 11] Inconsistent headline numbers: §3.6 reports U-Net + Semantic-Guided at 0.7183 mIoU / 0.2658 BF1 after constrained screening, while Table 11 and the abstract report 71.83% mIoU / 26.10% BF1. Please reconcile.
  2. [§2.4–2.5, Eqs. (3), (4), (11)] Three different objectives are introduced (L_stage1 with λ_ce L_ce, L_bco, and L_total with weighted CE), and Table 2 lists 'L_WCE+WDice' for Stage 1, which does not match Eq. (3). Please state unambiguously which loss is used in which stage and give the Stage-1 weight values.
  3. [§2.3–2.4] Key training details are missing: input resolution, diffusion timesteps and noise schedule, epoch counts for Stages 2–7, and the conditioning signal C for each refiner (Eq. 2). These are needed for reproducibility, especially given 'code available upon request' rather than a public release — a public repository with configs and seeds would substantially strengthen the paper.
  4. [§3.4, Table 6] The perturbation ranking is derived from a single model (DeepLabV3+ + DDPM) but the resulting augmentation policy is applied to all three hybrids and discussed generally. Either replicate the ranking on the other two models or justify why it transfers.
  5. [§3.7, Table 9] NWRD contains 100 images with a 0.85/0.05/0.10 split, so the test set is ~10 images; the zero-shot vs adapted numbers (0.5186 → 0.5892 mIoU) have very limited statistical power. Per-image variance or CIs should be reported, and similarly for CWFID (40 test images).
  6. [Figures and text] Figure 17 caption: 'selecrted' → 'selected'. 'Lovász'/'Lovasz' spelling is inconsistent. §3.9.1 contains a stray citation artifact ('remains challenging 9'). Table 2's tick/cross legend could note that 'global context' for SegFormer rows is marked ✗ for local detail, which is fine, but the column semantics would benefit from one clarifying sentence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical backbone–refiner screening and robustness protocol, not a derivation that equates claims to fitted identities.

full rationale

This paper is a staged empirical CV comparison (3×3 architectural screen → boundary-aware retrain → low-data → perturbation ranking → augmentation retrain → narrow HPO → few-shot transfer). Claims are measured mIoU/Dice/BF1 deltas on held validation or external sets, not first-principles predictions. Coarse→refined gains (Eqs. 1–2, Table 3) compare two model outputs under a fixed loss, not a quantity defined from the reported metric. Perturbation ranking (Stage 4) deliberately designs the Stage-5 curriculum; improved robustness on those shifts is expected engineering, not tautological ‘prediction.’ Top-K retention and constrained HPO on the source regime are ordinary model selection, not self-definitional reduction. No load-bearing uniqueness theorem or ansatz is imported via overlapping-author citation. Protocol concerns (class count, re-split, validation-only selection vs Wei et al.) affect external comparability, not circularity of the internal chain. Score 0; steps empty.

Assumptions & free parameters 6 free parameters · 6 assumptions · 2 invented entities

Load-bearing content is experimental protocol and fitted training choices, not new physical entities. Claims rest on standard segmentation/diffusion machinery, public datasets, ImageNet-pretrained encoders, composite losses with chosen weights, and the assumption that short-budget validation ranking selects generally good hybrids.

free parameters (6)
  • Stage-1 loss weights λ_ce, λ_dice = Not numerically specified beyond weighted CE+Dice
    Fixed shared screening objective; values treated as experimental settings after selection.
  • Boundary-constrained weights λ1, λ2, λ3 (Dice, boundary, Lovász) = e.g. Dice 1.0–1.5, boundary 1.0–1.25, Lovász 0.5–1.0
    Composite L_bco drives large BF1 gains; weights later retuned per model in Table 8.
  • Learning rates / batch sizes / weight decay / grad clip per hybrid = Table 8: LR 3e-5 or 5e-5; batch 4 or 8; etc.
    Constrained hyperparameter screen selects operating points that define the headline 71.83% mIoU checkpoint.
  • Boundary-F1 tolerance band = 1 pixel
    One-pixel tolerance makes BF1 strict and shapes ‘contour fidelity’ claims.
  • Screening budget (epochs, top-K retention) = 5 epochs; three hybrids retained
    5-epoch screen and top-K filter determine which architectures enter all later claims.
  • Augmentation curriculum phases and severities = Two-phase mild→severe policy (§2.8)
    Policy built from ranked drops; phase design is hand-structured around failure modes.
assumptions (6)
  • domain assumption Standard encoder–decoder and transformer segmentation backbones (U-Net, DeepLabV3+, SegFormer) provide valid coarse semantic localization for plant disease masks.
    Framework treats backbones as fixed coarse predictors (§2.3); supported by prior CV literature cited.
  • ad hoc to paper Diffusion modules act as structural refiners conditioned on image/coarse masks rather than standalone segmenters.
    Design restriction in §2.3; gains attributed to refinement deltas vs coarse baselines.
  • domain assumption PlantSegV3 train/val splits and nested low-data subsets represent realistic in-the-wild disease appearance sufficiently for robustness conclusions.
    Source of all screening and perturbation claims (§2.2, §3.8.1 redundancy analysis).
  • domain assumption ImageNet-pretrained encoders supply transferable visual priors that dominate low-label behavior.
    Table 10 shows ~0.20 mIoU gap vs random init at 10% labels; central to annotation-efficiency claim.
  • domain assumption Overlap metrics (mIoU/Dice), Boundary-F1, and pixel accuracy jointly measure stress-segmentation quality for phenotyping use.
    §2.11 metric suite; BF1 emphasized for thin fragmented lesions.
  • ad hoc to paper Short local-GPU training is an appropriate filter for ‘practical’ architectural conclusions.
    Stated design goal §§1–2.1; also the weakest ranking assumption.
invented entities (2)
  • Diffusion-guided hybrid segmentation framework (3×3 backbone–refiner design space)
    purpose: Organize coarse discriminative prediction plus diffusion refinement for plant stress masks under constrained training.
    Compositional system name for known parts; not a new physical entity. Independent evidence is only the paper’s own benchmarks.
  • Perturbation-derived two-phase augmentation policy
    purpose: Convert measured failure modes (grayscale, fog, dropout, shadow, …) into retraining curriculum.
    Procedural artifact of Stage 4→5; falsifiable by re-running sensitivity on other models/datasets, but not validated outside this study.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Perturbation-Aware Diffusion-Guided Hybrid Segmentation for Robust and Annotation-Efficient Plant Stress Phenotyping." pith.science (2026). https://pith.science/paper/YVPKOE6C

@misc{pith2026260723680,
  author       = {Pith},
  title        = {Pith review of: Perturbation-Aware Diffusion-Guided Hybrid Segmentation for Robust and Annotation-Efficient Plant Stress Phenotyping},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YVPKOE6C}},
  note         = {Machine review of arXiv:2607.23680}
}
read the original abstract

Semantic segmentation in agricultural imagery is often evaluated under in-domain protocols, yet practical deployment requires robustness to appearance perturbations, limited annotations, and cross domain shift. This paper presents a diffusion-guided hybrid segmentation framework in which U-Net, DeepLabV3+, and SegFormer backbones generate coarse masks that are refined by Denoising Diffusion Probabilistic Models (DDPM), latent diffusion, or semantic-guided diffusion. The framework is evaluated through a 3x3 architectural screening study on PlantSegV3, followed by boundary-constrained optimization, perturbation-guided retraining, low-data evaluation, constrained hyperparameter screening, and controlled cross-domain adaptation. On PlantSegV3, the best selected hybrid model achieves 71.83% refined mean Intersection-over-Union (mIoU) and 26.10% refined Boundary-F1, and the selected models remain stable under substantially reduced supervision, demonstrating strong annotation efficiency. Perturbation analysis identifies grayscale conversion, fog, coarse dropout, and shadow as the most disruptive appearance shifts, and the resulting augmentation policy substantially improves robustness during retraining. The adapted models further show effective transfer to external agricultural datasets under limited target supervision, indicating that diffusion refinement and boundary-aware optimization provide transferable structural priors. Overall, the results show that carefully matched backbone-refiner pairings, combined with perturbation-aware retraining, can improve structural delineation and robustness under realistic resource and distribution constraints.

Figures

Figures reproduced from arXiv: 2607.23680 by the authors.

Figure 1
Figure 1. Qualitative segmentation samples from the [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Qualitative segmentation samples from the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. Overview of the proposed diffusion-guided hybrid segmentation framework. The figure summarizes the coarse [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (11 more)
Figure 5
Figure 5. Figure 5: Architectures of the three selected hybrid segmentation configurations used in this study: (a) U-Net + semantic [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Qualitative comparison of coarse and diffusion [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Performance comparison of all screened hybrid [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Qualitative comparison of boundary-constrained [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Performance of the selected hybrid segmenta [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: Qualitative robustness assessment of the [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: Qualitative comparison of the baseline hybrid [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 12
Figure 12. Figure 12: Effect of increasing target-domain supervi [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]
Figure 14
Figure 14. Figure 14: Training dynamics of the selected hybrid con [PITH_FULL_IMAGE:figures/full_fig_p021_14.png]
Figure 15
Figure 15. Figure 15: Mean grayscale and RGB histogram profiles [PITH_FULL_IMAGE:figures/full_fig_p022_15.png]
Figure 17
Figure 17. Figure 17: Representative training samples selecrted from [PITH_FULL_IMAGE:figures/full_fig_p022_17.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

62 extracted references · 2 linked inside Pith

  1. [1]

    and Parashar, M

    Karthik, P. and Parashar, M. and Reka, S. S. and Rajamani, K. T. and Heinrich, M. P. , title =. Multimedia Tools and Applications , year =

  2. [2]

    Frontiers in Plant Science , year =

    Qin, Feng and Wang, Haiguang and Jiang, Qian and Wang, Hongli , title =. Frontiers in Plant Science , year =

  3. [3]

    Applied Sciences , volume =

    Song, Xuehua and Yan, Letian and Liu, Sihan and Gao, Tong and Han, Li and Jiang, Xiaoming and Jin, Hua and Zhu, Yi , title =. Applied Sciences , volume =. 2025 , doi =

    2025

  4. [4]

    and Mesquita, Ricardo and Simões, Maria P

    Assunção, Eduardo and Gaspar, Pedro D. and Mesquita, Ricardo and Simões, Maria P. and Alibabaei, Khadijeh and Veiros, André and Proença, Hugo , title =. Remote Sensing , volume =. 2022 , doi =

    2022

  5. [5]

    Electronics , volume =

    Yaseen, Muhammad and Ali, Maisam and Ali, Sikandar and Kim, Hee-Cheol , title =. Electronics , volume =. 2026 , doi =

    2026

  6. [6]

    Knowledge-Based Systems , volume =

    Zhiwei Wang and Defeng He and Li Zhao and Bo Liu and Yayu Zheng and Xiaoqin Zhang , title =. Knowledge-Based Systems , volume =. 2025 , doi =

    2025

  7. [7]

    Plant Phenomics , volume =

    Weizhen Liu and Guangyu Lan and Zhiwen Xing and Jiayu Tan and Jiale Liu and Qingzhong Liu and Dongzi Zhu and Xiaohui Yuan and Nanqing Dong , title =. Plant Phenomics , volume =. 2026 , doi =

    2026

  8. [8]

    , title =

    Alimisis, Panagiotis and Mademlis, Ioannis and Radoglou-Grammatikis, Panagiotis and Sarigiannidis, Panagiotis and Papadopoulos, Georgios Th. , title =. Artificial Intelligence Review , volume =. 2025 , doi =

    2025

Show all 62 references
  1. [9]

    ACM Computing Surveys , volume =

    Yang, Ling and Zhang, Zhilong and Song, Yang and Hong, Shenda and Xu, Runsheng and Zhao, Yue and Zhang, Wentao and Cui, Bin and Yang, Ming-Hsuan , title =. ACM Computing Surveys , volume =. 2023 , publisher =

    2023

  2. [10]

    Journal of King Saud University - Computer and Information Sciences , volume =

    Hongkui Jiang and Qiupu Chen and Rujing Wang and Jianming Du and Tianjiao Chen , title =. Journal of King Saud University - Computer and Information Sciences , volume =. 2024 , doi =

    2024

  3. [11]

    Anna Devi , title =

    Roshni Polly and E. Anna Devi , title =. Smart Agricultural Technology , volume =. 2024 , doi =

    2024

  4. [12]

    Green Technologies and Sustainability , volume =

    Indra Agustian and Igi Ardiyanto and Sunu Wibirama , title =. Green Technologies and Sustainability , volume =. 2026 , doi =

    2026

  5. [13]

    Information Fusion , volume =

    Aiqing Fang and Ying Li and Jiang Long and Xiaodong Wang and Yangming Guo , title =. Information Fusion , volume =. 2025 , doi =

    2025

  6. [14]

    arXiv preprint arXiv:2405.20443 , volume =

    Qi Zhang and Guohua Geng and Long-He Yan and Pengbo Zhou and Zhaodi Li and Kang Li and Qinglin Liu , title =. arXiv preprint arXiv:2405.20443 , volume =. 2024 , doi =

    2024 arXiv

  7. [15]

    and Le, Trung-Nghia and Do, Thanh-Toan and Do, Minh N

    Le, Minh-Quan and Nguyen, Tarn V. and Le, Trung-Nghia and Do, Thanh-Toan and Do, Minh N. and Tran, Minh-Triet , title =. Proceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence , pages =. 2024 , publisher =

    2024

  8. [16]

    Agronomy , volume =

    Shi, Xiaojie and Wang, Shaowei and Zhang, Bo and Ding, Xinbing and Qi, Peng and Qu, Huixing and Li, Ning and Wu, Jie and Yang, Huawei , title =. Agronomy , volume =. 2025 , doi =

    2025

  9. [17]

    Plant Methods , volume =

    Shafay, Muhammad and Hassan, Taimur and Owais, Muhammad and Hussain, Irfan and Khawaja, Sajid Gul and Seneviratne, Lakmal and Werghi, Naoufel , title =. Plant Methods , volume =. 2025 , doi =

    2025

  10. [18]

    Computers and Electronics in Agriculture , volume =

    Federico Magistri and Jan Weyler and Dario Gogoll and Philipp Lottes and Jens Behley and Nik Petrinic and Cyrill Stachniss , title =. Computers and Electronics in Agriculture , volume =. 2023 , doi =

    2023

  11. [19]

    Remote Sensing of Environment , volume =

    Ren Wei and Lin Yang and Xiang Li and Chenxu Zhu and Lei Zhang and Jie Wang and Jie Liu and Liming Zhu and Chen , title =. Remote Sensing of Environment , volume =. 2025 , doi =

    2025

  12. [20]

    Plant Phenomics , volume =

    Zihan Ge and Xijian Fan and Jingcheng Zhang and Shichao Jin , title =. Plant Phenomics , volume =. 2025 , doi =

    2025

  13. [21]

    and Zeevi, Tal and Payabvash, Seyedmehdi , title =

    Tran, Anh T. and Zeevi, Tal and Payabvash, Seyedmehdi , title =. BioMedInformatics , volume =. 2025 , doi =

    2025

  14. [22]

    Plant Phenomics , volume =

    Xinlu Wu and Xijian Fan and Peng Luo and Sruti Das Choudhury and Tardi Tjahjadi and Chunhua Hu , title =. Plant Phenomics , volume =. 2023 , doi =

    2023

  15. [23]

    Smart Agricultural Technology , volume =

    João Vitor de Andrade Porto and Arlinda Cantero Dorsa and Vanessa Aparecida de Moraes Weber and Karla Rejane de Andrade Porto and Hemerson Pistori , title =. Smart Agricultural Technology , volume =. 2023 , doi =

    2023

  16. [24]

    and Melloy, Paul and Huang, Zi , title =

    Wei, Tianqi and Chen, Zhi and Yu, Xin and Chapman, Scott C. and Melloy, Paul and Huang, Zi , title =. Scientific Data , volume =. 2026 , doi =

    2026

  17. [25]

    Journal of Rock Mechanics and Geotechnical Engineering , year =

    Shunzheng Jia and Yudi Tang and Yulin Wang and Xin Zhang and Joung Oh and Guangyao Si , title =. Journal of Rock Mechanics and Geotechnical Engineering , year =

  18. [26]

    arXiv preprint arXiv:2112.03126 , volume =

    Dmitry Baranchuk and Ivan Rubachev and Andrey Voynov and Valentin Khrulkov and Artem Babenko , title =. arXiv preprint arXiv:2112.03126 , volume =

    arXiv

  19. [27]

    Conference on Empirical Methods in Natural Language Processing , year =

    Vinodkumar Prabhakaran and Ben Hutchinson and Margaret Mitchell , title =. Conference on Empirical Methods in Natural Language Processing , year =

  20. [28]

    Knowledge-Based Systems , volume =

    Chunnan Wang and Hongzhi Wang and Chang Zhou and Hanxiao Chen , title =. Knowledge-Based Systems , volume =. 2021 , doi =

    2021

  21. [29]

    Artificial Intelligence Review , volume =

    Rafi, Tahmeed Tarek and Mahjabin, Radiah and Ghosh, Emon and Hoque, Mohammad Mehedi and Hassan, Mohammad Mehedi , title =. Artificial Intelligence Review , volume =. 2024 , doi =

    2024

  22. [30]

    Sensors , volume =

    Anwar, Hirra and Khan, Saad Ullah and Ghaffar, Muhammad Mohsin and Fayyaz, Muhammad and Khan, Muhammad Jawad and Weis, Christian and Wehn, Norbert and Shafait, Faisal , title =. Sensors , volume =. 2023 , doi =

    2023

  23. [31]

    A Crop/Weed Field Image Dataset for the Evaluation of Computer Vision Based Precision Agriculture Tasks , booktitle =

    Haug, Sebastian and Ostermann, J. A Crop/Weed Field Image Dataset for the Evaluation of Computer Vision Based Precision Agriculture Tasks , booktitle =. 2015 , doi =

    2015

  24. [32]

    Computer Vision -- ECCV 2018 , volume =

    Chen, Liang-Chieh and Zhu, Yukun and Papandreou, George and Schroff, Florian and Adam, Hartwig , title =. Computer Vision -- ECCV 2018 , volume =. 2018 , publisher =

    2018

  25. [33]

    and Luo, Ping , title =

    Xie, Enze and Wang, Wenhai and Yu, Zhiding and Anandkumar, Anima and Alvarez, Jose M. and Luo, Ping , title =. Proceedings of the 35th International Conference on Neural Information Processing Systems , year =

  26. [34]

    Medical Image Computing and Computer-Assisted Intervention -- MICCAI 2015 , pages =

    Ronneberger, Olaf and Fischer, Philipp and Brox, Thomas , title =. Medical Image Computing and Computer-Assisted Intervention -- MICCAI 2015 , pages =. 2015 , publisher =

    2015

  27. [35]

    Advances in Neural Information Processing Systems , volume =

    Ho, Jonathan and Jain, Ajay and Abbeel, Pieter , title =. Advances in Neural Information Processing Systems , volume =. 2020 , doi =

    2020

  28. [36]

    High-Resolution Image Synthesis with Latent Diffusion Models , booktitle =

    Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Patrick, Esser and Ommer, Bj. High-Resolution Image Synthesis with Latent Diffusion Models , booktitle =. 2022 , doi =

    2022

  29. [37]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages =

    Liu, Xihui and Park, Dong Huk and Azadi, Samaneh and Zhang, Gong and Chopikyan, Arman and Hu, Yuxiao and Shi, Humphrey and Rohrbach, Anna and Darrell, Trevor , title =. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages =. 2023 , doi =

    2023

  30. [38]

    Trends in Plant Science , volume =

    Asheesh Kumar Singh and Baskar Ganapathysubramanian and Soumik Sarkar and Arti Singh , title =. Trends in Plant Science , volume =. 2018 , doi =

    2018

  31. [39]

    and Cui, X

    He, Q. and Cui, X. and Fei, Q. and others , title =. Scientific Reports , volume =. 2025 , publisher =

    2025

  32. [40]

    Papadopoulos and Hoi Yeung , title =

    Christos E. Papadopoulos and Hoi Yeung , title =. Flow Measurement and Instrumentation , volume =. 2001 , doi =

    2001

  33. [41]

    and Persaud, D

    Li, K. and Persaud, D. and Choudhary, K. and others , title =. Nature Communications , volume =. 2023 , publisher =

    2023

  34. [42]

    Anisur Rahman and Swakkhar Shatabda and Md

    Mohaimenul Azam Khan Raiaan and Sadman Sakib and Nur Mohammad Fahad and Abdullah Al Mamun and Md. Anisur Rahman and Swakkhar Shatabda and Md. Saddam Hossain Mukta , title =. Decision Analytics Journal , volume =. 2024 , doi =

    2024

  35. [43]

    and Harder, B

    Stuckner, J. and Harder, B. and Smith, T. M. , title =. npj Computational Materials , volume =. 2022 , publisher =

    2022

  36. [44]

    Knowledge-Based Systems , volume =

    Luca Dini and Lucia Domenichelli and Dominique Brunato and Felice Dell’Orletta , title =. Knowledge-Based Systems , volume =. 2026 , doi =

    2026

  37. [45]

    and Ullah, K

    Hussain, A. and Ullah, K. and Afaq, M. and others , title =. Artificial Intelligence Review , volume =. 2026 , publisher =

    2026

  38. [46]

    2024 , doi =

    Peiliang Zhang and Yujia Tong and Chenghu Du and Chao Che and Yongjun Zhu , title =. 2024 , doi =

    2024

  39. [47]

    Proceedings of the 22nd ACM International Conference on Information & Knowledge Management , pages =

    To, Hien and Chiang, Kuorong and Shahabi, Cyrus , title =. Proceedings of the 22nd ACM International Conference on Information & Knowledge Management , pages =. 2013 , publisher =

    2013

  40. [48]

    Ferreira , title =

    Oliver Holub and Sérgio T. Ferreira , title =. Computer Physics Communications , volume =. 2006 , doi =

    2006

  41. [49]

    Statistical analysis of high-dimensional biomedical data: a gentle introduction to analytical goals, common approaches and challenges , journal =

    Rahnenf. Statistical analysis of high-dimensional biomedical data: a gentle introduction to analytical goals, common approaches and challenges , journal =. 2023 , publisher =

    2023

  42. [50]

    IEEE Signal Processing Magazine , volume =

    Minervini, Massimo and Scharr, Hanno and Tsaftaris, Sotirios , title =. IEEE Signal Processing Magazine , volume =. 2015 , doi =

    2015

  43. [51]

    Applied Sciences , volume =

    Seo, Suyoung , title =. Applied Sciences , volume =. 2021 , doi =

    2021

  44. [52]

    Frontiers in Plant Science , volume =

    Khan, Ameer Tamoor and Jensen, Signe Marie and Khan, Abdul Rehman and Li, Shuai , title =. Frontiers in Plant Science , volume =. 2023 , doi =

    2023

  45. [53]

    Horticulturae , volume =

    Gao, Bingpeng and Nie, Huishan and Du, Tiantian and Cai, Xin , title =. Horticulturae , volume =. 2026 , doi =

    2026

  46. [54]

    and Simegn, Gizeaddis Lamesgin , title =

    Degu, Mizanu Zelalem and Raymond, Confidence and Zhang, Dong and Saleh, Amal and Anazodo, Udunna C. and Simegn, Gizeaddis Lamesgin , title =. Medical Information Computing , pages =. 2025 , publisher =

    2025

  47. [55]

    Array , volume =

    Mobin Mohammadi and Kaveh Mollazade and Nasser Behroozi-Khazaei , title =. Array , volume =. 2025 , doi =

    2025

  48. [56]

    2021 14th International Conference on Human System Interaction (HSI) , pages =

    Elhassan Mohamed and Konstantinos Sirlantzis and Gareth Howells , title =. 2021 14th International Conference on Human System Interaction (HSI) , pages =. 2021 , doi =

    2021

  49. [57]

    Advances in Neural Networks – ISNN 2019: 16th International Symposium on Neural Networks , pages =

    Bokhovkin, Alexey and Burnaev, Evgeny , title =. Advances in Neural Networks – ISNN 2019: 16th International Symposium on Neural Networks , pages =. 2019 , publisher =

    2019

  50. [58]

    Computerized Medical Imaging and Graphics , volume =

    Michael Yeung and Evis Sala and Carola-Bibiane Schönlieb and Leonardo Rundo , title =. Computerized Medical Imaging and Graphics , volume =. 2022 , doi =

    2022

  51. [59]

    and Cordova-Esparza, D

    Terven, J. and Cordova-Esparza, D. M. and Romero-Gonz. A comprehensive survey of loss functions and metrics in deep learning , journal =. 2025 , publisher =

    2025

  52. [60]

    and Alaoui, S

    Fawakherji, M. and Alaoui, S. and Turjya, S. M. , title =. Archives of Computational Methods in Engineering , year =

  53. [61]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops , pages =

    Tavera, Antonio and Arnaudo, Edoardo and Masone, Carlo and Caputo, Barbara , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops , pages =. 2022 , doi =

    2022

  54. [62]

    2026 , eprint=

    Pixel-Precise Explainable Stress Indexing: A Semantic Segmentation Framework for Disease Severity Quantification in Field Crops , author=. 2026 , eprint=

    2026

Pith tools

Reviewed July 30, 2026 · model on record in the stance chip above.