Pith. sign in

REVIEW 4 major objections 5 minor 26 references

Multi-Modal Robust Enhancement for Coastal Water Segmentation: A Systematic HSV-Guided Framework

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper proposes a five-part HSV-guided loss for U-Net coastal water segmentation, claiming Lipschitz-based convergence guarantees, 84% lower training variance, and IoU of 0.9645.

desk verdict A plausible engineering recipe for HSV-guided water segmentation is buried under two invalid theorems and an underpowered single-region evaluation. read the letter →

arxiv 2509.08694 v1 pith:UY2GQGWL submitted 2025-09-10 cs.CV

Zhen Tian , Christos Anagnostopoulos , Qiyuan Wang , Zhiwei Gao This is my paper
classification cs.CV
keywords coastalwatersegmentationHSVcolorspaceU-NetsatelliteimageryrobusttrainingsemanticLipschitzcontinuitymulti-modalconstraints
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

The paper sets out to make coastal water segmentation from satellite images more stable and accurate by supervising a U-Net with HSV color-space priors and geometric constraints, rather than relying on RGB cross-entropy alone. It claims two theoretical results: every loss component is Lipschitz continuous, so the composite objective makes gradient descent converge to a stationary point at O(1/sqrt(T)), and the spectral-geometric priors shrink the effective hypothesis space. On Sentinel-2 imagery of a Scottish coastal region, the full framework reports an 84% reduction in training-loss variance, a final IoU of 0.9645, and cleaner coastlines than the vanilla U-Net baseline. The practical payoff would be segmentation models that are less sensitive to training noise and more trustworthy for monitoring shorelines over time. Both theorems are supported by proof sketches, and the authors state that full technical details are omitted due to space limitations.

What carries the argument

The load-bearing mechanism is the composite training objective L_Robust = lambda_CE L_CE + lambda_HSV L_HSV + lambda_Coast L_Coast + lambda_Conn L_Conn + lambda_Sea L_Sea, whose claimed regularity (each term Lipschitz continuous) underwrites Theorem 1's convergence guarantee. The main inductive bias is the HSV-guided water likelihood P_HSV(i,j) = sigma(alpha_H H_ij + alpha_S S_ij + alpha_V V_ij + beta), a learned probability that anchors supervision to the spectral signature of water. Geometric terms enforce smooth coastline gradients, penalize disconnected water regions per column, and suppress local variance inside detected sea areas.

What would settle it

Evaluate Eq. (8) on two masks M1 and M2 that differ only by a tiny positive bump connecting two previously separate water blobs in one column: the connected-component count drops by 1 while the difference norm can be made arbitrarily small, so the ratio |L_Conn(M1)-L_Conn(M2)|/||M1-M2|| is unbounded, directly contradicting the Lipschitz premise used in Theorem 1.

Watch

Extended reading notes

Core claim

On its own terms, the paper's core discovery is that water bodies have stable, low-saturation HSV signatures, and that encoding this prior through an additional supervision loss—together with penalties on coastline jaggedness, disconnected water columns, and intra-body variance—lets a standard U-Net train more reliably and segment coastlines more faithfully. The authors claim to prove (Theorem 1) that the composite objective is Lipschitz continuous with constant L = max_i lambda_i L_i, yielding O(1/sqrt(T)) convergence of gradient descent, and (Theorem 2) that the priors reduce the effective hypothesis space by a factor proportional to the correlation between HSV features and water labels. E

Load-bearing premise

The convergence guarantee in Theorem 1 assumes every component of the composite loss is Lipschitz continuous, including the connectivity term that counts connected components in a binary sequence derived from the predicted mask, and that the composite Lipschitz constant is a max of weighted constants rather than a sum.

Editorial extensions

If this is right

  • If the convergence guarantee holds, practitioners can train coastal-water U-Nets with the composite loss and expect gradient descent to reach a stationary point at O(1/sqrt(T)), reducing the need for hand-tuned training schedules.
  • The 84% reduction in training-loss variance implies more reproducible segmentation outputs across epochs, which matters for time-series coastal monitoring where consistency across revisits is as important as raw accuracy.
  • The reported IoU gain (0.9645 versus a vanilla U-Net baseline) suggests that the constraints improve boundary accuracy without architectural changes, so the method can be applied to existing U-Net deployments.
  • The framework keeps computational cost low while combining five loss terms, making it practical for large satellite-image archives.
  • The high ablation influence of HSV supervision (0.85) suggests that color-space supervision, not post-processing, is the primary source of the improvement.

Reading between the lines

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

  • The convergence theorem's validity turns on whether the connected-component count in Eq. (8) is continuous in the predicted mask; replacing it with a differentiable proxy (such as a total-variation or pairwise-affinity penalty) would make the Lipschitz claim testable and likely recoverable.
  • Because HSV supervision is end-to-end trainable through learned coefficients alpha_H, alpha_S, alpha_V, the approach could plausibly transfer to other water bodies—lakes, rivers, floods—where hue and saturation signatures differ, provided the coefficients are re-estimated per environment.
  • A direct experiment would be to train the same U-Net with only L_HSV and compare training variance to the full five-term objective; if most of the 84% variance reduction persists, the geometric terms are not load-bearing for stability.
  • The reported influence scores (HSV 0.85, coastline 0.79, morphology 0.74, sea cleanup 0.72, connectivity 0.68) suggest that the components are not independent contributors; a future ablation could test whether adding the weaker terms in any order improves on HSV supervision alone.
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 / 5 minor

Summary. The paper proposes Robust U-Net, a U-Net variant for coastal water segmentation that adds HSV-guided supervision, coastline smoothness and connectivity losses, a sea-cleanup regularizer, and post-processing. The abstract and introduction claim two theoretical results: Theorem 1, that the composite loss is Lipschitz and hence gradient descent converges at O(1/sqrt(T)), and Theorem 2, that spectral/geometric constraints shrink the effective hypothesis space. Experiments on Sentinel-2 coastal imagery report lower training variance and marginally higher IoU/F1 than DeepLabV3+ and YOLO-SEG, with an ablation figure attributing influence scores to each component. A repository link for code is provided.

Significance. If the theoretical guarantees were sound, the paper would offer a novel and useful bridge between domain-specific HSV priors, geometric regularization, and optimization theory for remote-sensing segmentation. The empirical recipe is potentially useful engineering. However, the central theoretical claims are not established: Theorem 1 is based on a false Lipschitz premise and an incorrect composite-Lipschitz expression, and Theorem 2 is a definition rather than a proof. The reported improvements over DeepLabV3+ are within error bars, and the comparison is confounded by additional architectural changes. The paper's significance as submitted is therefore limited to a heuristic combination of losses, and the claimed guarantees should not be relied on.

major comments (4)
  1. [Section 3.2, Eq. (1) and Eq. (10)] Theorem 1 is not valid as stated. First, for a sum of Lipschitz losses the composite constant should be at most the sum Σ_i λ_i L_i, not L = max_i λ_i L_i; the displayed inequality is not generally true. Second, the premise that every component is Lipschitz fails for L_Conn in Eq. (8): ConnectedRegions(M[:,x]) counts connected components of a binary 1D sequence obtained from a continuous mask. Such a count is piecewise constant and discontinuous, so L_Conn is not Lipschitz. Since L_Conn appears in Eq. (10), Theorem 1 does not apply to L_Robust. The proof sketch also cites [20], which requires Lipschitz-continuous gradients (smoothness), a stronger condition than Lipschitz continuity of the losses. Thus the advertised 'guarantee of reliable optimization' is unsupported.
  2. [Section 3.2, Eq. (2)] Theorem 2 is a definition, not a theorem. It introduces ρ as the correlation between HSV features and ground-truth labels and then asserts H_Reduced ≈ ρ · H_Full, without defining the hypothesis spaces, the multiplication operation, or a measure of cardinality. The claimed reduction is stipulated by the definition of ρ, and the reference to prior-domain-knowledge work does not provide a formal derivation. This is load-bearing because the introduction and contribution list use Theorem 2 to claim that the proposed constraints 'shrink the effective hypothesis space.'
  3. [Section 4.4, Table 1] The headline comparison does not support the claimed superiority. Robust U-Net achieves IoU 0.9645±0.003 vs DeepLabV3+ 0.9639±0.005, F1 0.9819±0.002 vs 0.9816±0.003, and Acc 0.9810±0.002 vs 0.9806±0.003. These differences are well within the reported standard deviations, so the statement that the method 'clearly outperforms' DeepLabV3+ is not backed by the table. Moreover, Sec. 4.4 says Robust U-Net was 'further enhanced' with dropout, multi-scale dilated convolutions, attention, and Kaiming initialization; the comparison therefore conflates architectural changes with the proposed robust losses, and the ablation in Fig. 1 does not isolate this confound.
  4. [Section 4.2, Figs. 1–2] The 'influence score' (e.g., 0.85 for HSV supervision) is not defined, and no statistical protocol is described for its computation. The claimed '84% variance reduction' is based on two aggregated numbers (0.0079 vs 0.0012) without specifying what is being averaged, how many seeds were used, or whether the variance is over runs, pixels, or epochs. These quantities are therefore not independently verifiable from the manuscript.
minor comments (5)
  1. [Section 3.3, Eq. (3)] Calling the HSV likelihood 'physics-informed' is overstated because α_H, α_S, α_V, β are learnable parameters fitted to the ground-truth labels; the name suggests a more principled physical prior than is provided.
  2. [Section 4.1] The input is described as RGB in Section 3.1, but the dataset uses NIR-Red-Green band combination. The color-space conversion path from this three-channel input to RGB and then HSV should be clarified.
  3. [Section 3.4, Eq. (8)] The binarization threshold applied to M[:,x] before ConnectedRegions is not specified, nor is the value of MaxRegions. These details are necessary to reproduce the connectivity loss.
  4. [Section 4.1] There is no independent test set; all reported numbers appear to be on a 20% validation split. This should be stated explicitly, and results should ideally be reported on a held-out test set.
  5. [Section 3.2] The proof sketches are said to be omitted 'due to space limitations.' For a journal submission, the full proofs are not optional when the theorems are central to the paper's claims.

Circularity Check

1 steps flagged · score 6.0 of 10

Theorem 2's hypothesis-space reduction is a definitional tautology; empirical results are self-contained.

  1. self definitional [Section 3.2, Theorem 2 statement and Eq. (2)]
    "The combined influence of spectral priors and geometric regularizers reduces the effective hypothesis space according to HReduced ≈ρ· HFull (2) where ρ∝Corr(HSV features,true water labels)quantifies the correlation between HSV-based indicators and ground truth water labels"

    The theorem's claimed reduction is not derived. Eq. (2) defines H_Reduced as ρ times H_Full, and ρ is itself defined as the correlation between HSV features and ground-truth labels. The proof sketch only states that adding constraints K gives a subset H' ⊆ H, which does not imply any quantitative scaling by ρ. The 'result' is therefore true by stipulation: the reduction factor is exactly the correlation used to define it. This is a self-definitional tautology rather than a proof, so the claim that spectral–geometric constraints shrink the hypothesis space is equivalent to its own definition.

full rationale

The paper's empirical evaluation—variance reduction, IoU comparisons, ablation influence scores—is self-contained and does not reduce to its inputs; I find no self-citation chain. The one genuine circular step is Theorem 2: Eq. (2) defines the reduced hypothesis space as proportional to the correlation between HSV features and ground-truth labels, and the proof sketch only asserts that constraints yield a subset, which does not imply that quantitative scaling. Thus the claimed 'hypothesis space reduction' is a stipulated definition, not a derived result. Separately, Theorem 1 is not circular but is a correctness risk: the paper itself admits only proof sketches are given, and the connectivity loss L_Conn in Eq. (8) counts connected components of a binarized sequence and is discontinuous, so the Lipschitz premise fails; the composite Lipschitz constant in Eq. (1) is also stated as a max rather than a sum. These are mathematical errors, not cases where a prediction reduces to a fit. Score reflects the one definitional theoretical claim.

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

The central claim rests on several fitted coefficients and hand-set hyperparameters: the HSV water-likelihood coefficients (learned), the loss weights λ, the bandwidth σ, MaxRegions, and the sea-pixel area threshold. It also relies on domain assumptions about HSV stability across water types and on ad hoc postulates in Theorem 2. The theorems assume away the non-smoothness of the connectivity loss and define hypothesis-space reduction via the same correlation that is used to justify it.

free parameters (5)
  • α_H, α_S, α_V, β (HSV water-likelihood coefficients) = learned during training, not reported
    Eq. 3 defines P_HSV as a sigmoid over a linear combination of HSV channels with learnable coefficients. These are fitted to the ground-truth labels, so the 'physics-informed' prior is actually a learned classifier.
  • Loss weights λ_CE, λ_HSV, λ_Coast, λ_Conn, λ_Sea = not reported
    Eq. 10 combines five losses; the weights are hand-chosen or tuned on the validation set. No sensitivity analysis or values are given, and the 'influence scores' depend on these weights.
  • Bandwidth σ in the confidence weight w_ij = not reported
    Eq. 5 uses a Gaussian weighting with bandwidth σ; its choice is critical to the HSV loss but is not specified or justified.
  • MaxRegions normalization in L_Conn = not reported
    Eq. 8 divides by MaxRegions to bound the connectivity loss; the paper never defines it. It must be set somehow.
  • Area threshold for sea pixels S in L_Sea = not reported
    Eq. 9 uses connected component analysis with area thresholding to define S, but the threshold is not given.
assumptions (4)
  • standard math Cross-entropy and L2 losses are Lipschitz on the bounded probability mask domain [0,1]
    Assumed by Theorem 1; reasonable for smooth losses on a compact domain.
  • domain assumption Water bodies show stable, low-saturation HSV signatures across lighting and water types
    Section 1 states this as the key idea. It is an empirical claim about the data and may fail for turbid or sediment-laden coastal waters.
  • ad hoc to paper ConnectedRegions on the continuous mask is differentiable/Lipschitz
    Necessary for Theorem 1 to cover L_Conn (Eq. 8), but counting connected components of a binarized column is piecewise constant and discontinuous, so the axiom is false.
  • ad hoc to paper H_Reduced ≈ ρ·H_Full with ρ ∝ Corr(HSV, labels)
    Theorem 2 postulates this relationship without derivation; it is a definition of the reduction scale, not a proved bound.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Modal Robust Enhancement for Coastal Water Segmentation: A Systematic HSV-Guided Framework." pith.science (2026). https://pith.science/paper/UY2GQGWL

@misc{pith2026250908694,
  author       = {Pith},
  title        = {Pith review of: Multi-Modal Robust Enhancement for Coastal Water Segmentation: A Systematic HSV-Guided Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UY2GQGWL}},
  note         = {Machine review of arXiv:2509.08694}
}
read the original abstract

Coastal water segmentation from satellite imagery presents unique challenges due to complex spectral characteristics and irregular boundary patterns. Traditional RGB-based approaches often suffer from training instability and poor generalization in diverse maritime environments. This paper introduces a systematic robust enhancement framework, referred to as Robust U-Net, that leverages HSV color space supervision and multi-modal constraints for improved coastal water segmentation. Our approach integrates five synergistic components: HSV-guided color supervision, gradient-based coastline optimization, morphological post-processing, sea area cleanup, and connectivity control. Through comprehensive ablation studies, we demonstrate that HSV supervision provides the highest impact (0.85 influence score), while the complete framework achieves superior training stability (84\% variance reduction) and enhanced segmentation quality. Our method shows consistent improvements across multiple evaluation metrics while maintaining computational efficiency. For reproducibility, our training configurations and code are available here: https://github.com/UofgCoastline/ICASSP-2026-Robust-Unet.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 1 linked inside Pith

  1. [20]

    Mim-unet: An efficient building image segmentation network integrating state space models,

    Dong Liu et al., “Mim-unet: An efficient building image segmentation network integrating state space models,” Alexandria Engineering Journal, vol. 120, pp. 648–656, 2025

    2025

  2. [1]

    Traditional RGB-based methods struggle in maritime scenes due to spectral variability, atmospheric effects, and complex shorelines [2, 3, 4]

    INTRODUCTION Coastal water segmentation from satellite and aerial im- agery is a core task in remote sensing, with applications in environmental monitoring, climate studies, and coastal management [1]. Traditional RGB-based methods struggle in maritime scenes due to spectral variability, atmospheric effects, and complex shorelines [2, 3, 4]. Deep learning...

    2026

  3. [2]

    Deep Learning for Water Segmentation Traditional water segmentation methods relied on spectral in- dices such as NDWI and MNDWI [13, 14]

    RELATED WORK 2.1. Deep Learning for Water Segmentation Traditional water segmentation methods relied on spectral in- dices such as NDWI and MNDWI [13, 14]. With the ad- vent of deep learning, convolutional neural networks have become the dominant approach. U-Net [15] and its variants have shown particular success in biomedical and remote sens- ing applica...

    arXiv 2025

  4. [3]

    METHODOLOGY We introduce a novel framework, coined Robust U-Net, for robust coastline delineation that integrates physics-informed supervision, geometric regularization, and regional consis- tency enforcement. Robust U-Net is grounded in two fun- damental theoretical properties that ensure convergence guar- antees and hypothesis space reduction, providing...

  5. [4]

    Datasets Experiments are conducted on coastal imagery from St An- drews region, Scotland, acquired via Sentinel-2 satellite im- ages from May 2017 to May 2025

    EXPERIMENTAL EV ALUATION 4.1. Datasets Experiments are conducted on coastal imagery from St An- drews region, Scotland, acquired via Sentinel-2 satellite im- ages from May 2017 to May 2025. The dataset comprises 954 high-resolution images with pixel-level water segmenta- tion annotations. Images utilize NIR-Red-Green band combi- nation for enhanced water-...

    2017

  6. [5]

    Theory and experiments show improved stability, faster con- vergence, and higher accuracy over baselines

    CONCLUSIONS We introduce a robust framework for coastal water segmen- tation that integrates HSV-guided supervision, geometric and morphological constraints, and sea cleanup regularization. Theory and experiments show improved stability, faster con- vergence, and higher accuracy over baselines. Robust U-Net achieves the best performance (IoU 0.9645, F1 0....

  7. [6]

    Vedgesat: An automated, open-source toolkit for coastal change monitoring using satellite-derived vegetation edges,

    Freya ME Muir et al., “Vedgesat: An automated, open-source toolkit for coastal change monitoring using satellite-derived vegetation edges,”Earth Surface Pro- cesses and Landforms, vol. 49, no. 8, pp. 2405–2423, 2024

    2024

  8. [7]

    Differentiating biological colours with few and many sensors: spectral reconstruction with rgb and hyperspectral cameras,

    Jair E Garcia et al., “Differentiating biological colours with few and many sensors: spectral reconstruction with rgb and hyperspectral cameras,”PLoS One, vol. 10, no. 5, pp. e0125817, 2015

    2015

Show all 26 references
  1. [8]

    Living up to the hype of hy- perspectral aquatic remote sensing: science, resources and outlook,

    Heidi M Dierssen et al., “Living up to the hype of hy- perspectral aquatic remote sensing: science, resources and outlook,”Frontiers in Environmental Science, vol. 9, pp. 649528, 2021

    2021

  2. [9]

    Assessing landsat-8 atmospheric correc- tion schemes in low to moderate turbidity waters from a global perspective,

    Nanyang Yan, Zhen Sun, Wei Huang, Zhao Jun, and Shaojie Sun, “Assessing landsat-8 atmospheric correc- tion schemes in low to moderate turbidity waters from a global perspective,”International Journal of Digital Earth, vol. 16, no. 1, pp. 66–92, 2023

    2023

  3. [10]

    Unet-like remote sensing change de- tection: A review of current models and research direc- tions,

    Chen Wu et al., “Unet-like remote sensing change de- tection: A review of current models and research direc- tions,”IEEE Geoscience and Remote Sensing Maga- zine, 2024

    2024

  4. [11]

    U-net-based seman- tic classification for flood extent extraction using sar im- agery and gee platform: A case study for 2019 central us flooding,

    Zhouyayan Li and Ibrahim Demir, “U-net-based seman- tic classification for flood extent extraction using sar im- agery and gee platform: A case study for 2019 central us flooding,”Science of The Total Environment, vol. 869, pp. 161757, 2023

    2019

  5. [12]

    Improved unet-based shoreline de- tection method in real time for unmanned surface vehi- cle,

    Jiansen Zhao et al., “Improved unet-based shoreline de- tection method in real time for unmanned surface vehi- cle,”Journal of Marine Science and Engineering, vol. 11, no. 5, pp. 1049, 2023

    2023

  6. [13]

    A hybrid deep learning approach for accurate water body segmentation in satel- lite imagery,

    Mohammed Attya, OM Abo-Seida, HM Abdulkader, and Amgad M Mohammed, “A hybrid deep learning approach for accurate water body segmentation in satel- lite imagery,”Earth Science Informatics, vol. 18, no. 2, pp. 418, 2025

    2025

  7. [14]

    Bdcn unet: Advanced shoreline extraction techniques integrating deep learn- ing,

    Amira S Mahmoud, Sayed A Mohamed, Ashraf K Helmy, and Ayman H Nasr, “Bdcn unet: Advanced shoreline extraction techniques integrating deep learn- ing,”Earth Science Informatics, vol. 18, no. 2, pp. 187, 2025

    2025

  8. [15]

    Semantic segmentation framework for atoll satel- lite imagery: An in-depth exploration using unet vari- ants and segmentation gym,

    Ray Wang, Tahiya Chowdhury, and Alejandra C Or- tiz, “Semantic segmentation framework for atoll satel- lite imagery: An in-depth exploration using unet vari- ants and segmentation gym,”Applied Computing and Geosciences, vol. 25, pp. 100217, 2025

    2025

  9. [16]

    Deep and machine learning image classification of coastal wetlands using unpiloted aircraft system multispectral images and lidar datasets,

    Ali Gonzalez-Perez et al., “Deep and machine learning image classification of coastal wetlands using unpiloted aircraft system multispectral images and lidar datasets,” Remote Sensing, vol. 14, no. 16, pp. 3937, 2022

    2022

  10. [17]

    Hg2former: Hsv-gamma guided transformers for efficient underwater image enhancement,

    Yuhao Qing, Liquan Shen, Zhijun Fang, and Yueying Wang, “Hg2former: Hsv-gamma guided transformers for efficient underwater image enhancement,”IEEE Journal of Oceanic Engineering, 2025

    2025

  11. [18]

    Drought monitoring of large lakes in iraq using remote sensing images and normalized difference water index (ndwi),

    Mohammed R Mahmood, Baydaa Ismail Abrahem, Huda J Jumaah, Hayder A Alalwan, and Malik M Mo- hammed, “Drought monitoring of large lakes in iraq using remote sensing images and normalized difference water index (ndwi),”Results in Engineering, vol. 25, pp. 103854, 2025

    2025

  12. [19]

    A new spectral index for mapping water surfaces in urban contexts,

    Eduardo Felix Justiniano et al., “A new spectral index for mapping water surfaces in urban contexts,”IEEE Journal of Selected Topics in Applied Earth Observa- tions and Remote Sensing, 2025

    2025

  13. [21]

    Dsia u-net: deep shallow interaction with at- tention mechanism unet for remote sensing satellite im- ages,

    Naga Surekha Jonnala, Renuka Chowdary Bheemana, Krishna Prakash, Shonak Bansal, Arpit Jain, Vaibhav Pandey, Mohammad Rashed Iqbal Faruque, and KS Al- Mugren, “Dsia u-net: deep shallow interaction with at- tention mechanism unet for remote sensing satellite im- ages,”Scientific...

    2025

  14. [22]

    Msfm-unet: enhancing medical im- age segmentation with multi-scale and multi-view fre- quency fusion,

    Qiang Gao et al., “Msfm-unet: enhancing medical im- age segmentation with multi-scale and multi-view fre- quency fusion,”Pattern Analysis and Applications, vol. 28, no. 1, pp. 17, 2025

    2025

  15. [23]

    Hyperspectral video tracking with spectral-spatial fusion and memory enhancement,

    Yuzeng Chen et al., “Hyperspectral video tracking with spectral-spatial fusion and memory enhancement,” IEEE Transactions on Image Processing, 2025

    2025

  16. [24]

    Corn leaf diseases classifi- cation using cnn with glcm, hsv, and l* a* b* features,

    Putri Fausyah Johari et al., “Corn leaf diseases classifi- cation using cnn with glcm, hsv, and l* a* b* features,” Jurnal Teknik Informatika (Jutif), vol. 6, no. 2, pp. 709– 722, 2025

    2025

  17. [25]

    Stochastic first- and zeroth-order methods for nonconvex stochastic pro- gramming,

    Saeed Ghadimi and Guanghui Lan, “Stochastic first- and zeroth-order methods for nonconvex stochastic pro- gramming,”SIAM Journal on Optimization, vol. 23, no. 4, pp. 2341–2368, 2013

    2013

  18. [26]

    Incorporating prior domain knowledge into in- ductive machine learning,

    Ting Yu, Tony Jan, Simeon Simoff, and John Deben- ham, “Incorporating prior domain knowledge into in- ductive machine learning,”International Journal of Ar- tificial Intelligence Tools, vol. 16, no. 03, pp. 341–357, 2007

    2007

Pith tools

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