Pith. sign in

REVIEW 3 major objections 6 minor 21 references

Fruit-HSNet: A Machine Learning Approach for Hyperspectral Image-Based Fruit Ripeness Prediction

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Fruit-HSNet predicts fruit ripeness from hyperspectral images at 70.73% accuracy, a 12-point gain over the previous best model.

desk verdict A simple, sensible HSI classifier whose headline 12-point gain is unverified: no documented split, tiny per-config test sets, and Eq. (3) contradicts the text. read the letter →

arxiv 2608.01202 v1 pith:JI6WN6T5 submitted 2026-08-02 cs.CV cs.LG

Ahmed Baha Ben Jmaa , Faten Chaieb , Anna Fabijańska This is my paper · ORCID
classification cs.CVcs.LG
keywords FruitripenesspredictionHyperspectralimagingFouriertransformSpatio-spectralfeaturesDeepHSdatasetLearnablefeaturefusionSmartagriculture
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 introduces Fruit-HSNet, a compact architecture for classifying fruit ripeness from hyperspectral images. It extracts two complementary descriptors: the Fourier magnitude spectrum of each spectral band, capturing periodic skin textures, and the central pixel's full spectral signature, capturing chemical ripeness cues. These are fused with learnable weights and fed to a small fully connected classifier. Evaluated on the DeepHS Fruit benchmark, covering five fruits and three cameras, the method reports 70.73% overall accuracy, roughly 12 percentage points above the prior best DeepHS-Net. The ablation study attributes the gain to each component: removing the Fourier branch, the central-pixel branch, or the learnable fusion drops accuracy by 22, 12, and 10 points respectively.

What carries the argument

The dual-branch spatio-spectral feature extractor with learnable fusion: for each spectral channel $\lambda$, the 2D Fourier transform of the spatial image is computed and its magnitude averaged to a per-band scalar $m_\lambda$; in parallel, the central pixel's reflectance at each band gives $s_\lambda$. The two $B$-dimensional vectors are weighted by learnable vectors $w_1, w_2 \in \mathbb{R}^B$, combined, and concatenated into a $2B$-dimensional feature vector fed to a two-layer MLP classifier. The Fourier branch exposes periodic skin texture changes; the central pixel captures the chemically informative region; the learnable weights let the model reweight the two descriptors per configura

What would settle it

Re-run Fruit-HSNet and DeepHS-Net with a fruit-wise split (all images of a given fruit specimen assigned to one partition) using repeated k-fold cross-validation, and report the mean and standard deviation of overall accuracy across the 30 configurations. If the mean gap drops below roughly 5 points or overlaps with uncertainty, the claimed 12-point SOTA improvement would not hold.

Watch

Extended reading notes

Core claim

The central claim is that a deliberately simple, task-specific architecture can outperform much larger and more generic hyperspectral classifiers on fruit ripeness, provided the features are chosen to match the physics of ripening. Fruit-HSNet treats each hyperspectral image as a stack of spatial maps; it computes the 2D Fourier transform of each band and averages the magnitudes to form a spectral-texture descriptor, and it reads the reflectance spectrum at the image center as the chemical/ripeness descriptor. Learnable vectors weight the two descriptors before concatenation, letting the model adapt to fruit type and camera. On the DeepHS Fruit dataset, consisting of 2,706 labeled images of

Load-bearing premise

The 70.73% result depends on the data-splitting protocol adopted from Frank et al. (2023): if training and test images can contain the same fruit specimen, or if some per-configuration test sets hold only a handful of images, the reported 12-point gain could shrink.

Editorial extensions

If this is right

  • If the 70.73% figure holds under the adopted split, Fruit-HSNet becomes the best published result on the DeepHS Fruit benchmark and the first to push average accuracy past 70%.
  • The roughly 12-point gain over DeepHS-Net suggests that generic hyperspectral classifiers (2D/3D CNNs, transformers) are not exploiting ripeness-relevant structure, and that spectral-texture plus a single-point spectrum is a strong inductive bias for this task.
  • Camera-specific results imply that short-wave infrared cameras (920–1730 nm) carry enough information for ripeness discrimination: ripeness accuracy on Camera 3 reaches 88.89%.
  • The architecture is small and simple, so it is plausible to deploy in real-time or IoT settings, which the paper lists as future work.

Reading between the lines

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

  • The central-pixel spectral signature is a surprisingly strong spatial feature; a natural extension is to test whether multiple fiducial points (e.g., a grid or the fruit's centroid) generalize better when fruits are not perfectly centered in the frame.
  • Fourier magnitude averaging discards phase information, which may carry texture orientation cues; a phase-sensitive descriptor could improve discrimination between firmness and ripeness on the same fruit.
  • Because the reported gain is concentrated in mango and Camera 3, a testable prediction is that the Fourier branch matters most for fruits with pronounced skin texture changes, while the central-pixel branch matters most for fruits whose ripening is internal, like avocado.
  • The absence of error bars and the small per-configuration test sets suggest that re-running the comparison with repeated stratified splits would clarify whether the 12-point accuracy gap is statistically robust.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes Fruit-HSNet, a hyperspectral image classification architecture for fruit ripeness prediction. It extracts spectral features via the average Fourier-transform magnitude per band and spatial features via the central-pixel spectral signature, combines them with learnable weights, and classifies with a fully connected network. The method is evaluated on the DeepHS Fruit dataset across 30 configurations (five fruits, three cameras, three ripeness-related categories), reporting an overall accuracy of 70.73%, a 12.45-point gain over DeepHS-Net, together with fruit-, camera-, and component-level ablations. The paper claims a new state of the art on this dataset.

Significance. If the reported results are reproducible, Fruit-HSNet would be a practically relevant contribution to hyperspectral fruit ripeness prediction: the architecture is simple, the Fourier-plus-central-pixel feature design is clearly motivated, and the paper includes a systematic comparison against many baseline and state-of-the-art methods as well as an ablation study. These are genuine strengths. However, the central SOTA claim currently rests on an underspecified evaluation protocol: the data split is not described, no error bars or repeated runs are given, and some per-configuration metrics are almost certainly based on very small test sets. In addition, Eq. (3) is inconsistent with the described concatenation-based fusion, so it is not clear which architecture actually produced the numbers. The idea is promising, but the empirical claim needs to be substantiated with transparent and statistically valid evaluation.

major comments (3)
  1. [§3.1, Tables 1, 6, 7] The evaluation protocol is underspecified. The paper only states that the standard data splitting and preprocessing procedures of Frank et al. (2023) were followed, without reporting train/test sizes, whether the split is by image or by fruit specimen, class proportions, or the number of repetitions. This matters because several reported values, e.g., 88.89% for Camera 3 avocado ripeness and 100% for Camera 2 avocado firmness (Tables 6 and 7), are consistent with test sets of 9 or 27 images, where the binomial standard error is several percentage points. The headline 12.45-point improvement over DeepHS-Net may therefore be within sampling variability. Please report exact per-configuration test sizes, the split protocol (including any fruit-level separation), and confidence intervals or a paired significance test.
  2. [§2.2, Eq. (3) and §3.1] Eq. (3) defines f = w1*m + w2*s, which is a weighted sum producing a B-dimensional vector. The text immediately before Eq. (3) says the weighted features are 'then concatenated to form a feature vector f', and §3.1 specifies fully connected input dimensions [2λ, ...], implying concatenation. This is a direct inconsistency: a B-dim weighted sum cannot feed a 2λ-dim first layer. The reader cannot determine whether the evaluated architecture uses concatenation, a weighted sum, or a combination. Please correct Eq. (3) to, e.g., f = [w1⊙m; w2⊙s] for concatenation with per-band learnable weights, or adjust the classifier dimensions and all related descriptions.
  3. [§3.2.1, Table 1] The comparison with baselines is not fully fair as presented. Table 1 lists accuracy values for many methods 'obtained from (Frank et al., 2023)', while Fruit-HSNet is evaluated by the authors. It is not stated explicitly that every baseline was trained and tested under exactly the same split and preprocessing pipeline as Fruit-HSNet. If the deep learning baselines were evaluated under a different split, or if the Frank et al. numbers are the best over different protocols, the claimed 12.45-point SOTA advantage may be an artifact of protocol differences. Please either re-run all competing methods under the identical evaluation protocol or clearly state the source of each number and confirm split equivalence.
minor comments (6)
  1. [Abstract / §3.2.1] The phrase 'improvements of 12%' is ambiguous. The actual difference between 70.73% and 58.28% is 12.45 percentage points (a relative improvement of about 21.4%). Please state the comparison clearly as percentage points or relative gain.
  2. [§3.2.2, Table 3] The text refers to 'fruitiness' but the dataset categories are ripeness, firmness, and sweetness. Please fix the terminology. Similarly, §2.1 contains typos such as 'categoriei'.
  3. [§3.2.4 and Table 5] The paper itself acknowledges that Camera 3 sweetness contains only kiwi fruit, yet Table 5 reports a camera-level 'Sweetness' row for Camera 3. This should be labeled as a single-fruit configuration, not a general camera result, to avoid overgeneralization.
  4. [§3.1 and Figures 2–3] The figures and Figure 2 in particular would benefit from error bars or confidence intervals. Without them, the visual comparisons overstate the precision of the reported accuracy values.
  5. [§3.2.1] The global accuracy is described as an average across 30 datasets. Please clarify whether this is a macro-average over configurations or a pooled accuracy over all test samples, and how class imbalance is handled.
  6. [Reproducibility] The paper does not mention code, model checkpoints, or hyperparameter search details (e.g., how the variable number of epochs per configuration was selected). Releasing code would strengthen the reproducibility of the claimed SOTA result.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the 70.73% result is a trained-model test accuracy, not a fitted quantity renamed as a prediction.

full rationale

Fruit-HSNet is an empirical machine-learning paper: the headline 70.73% accuracy is a measured test-set outcome of training a classifier, not the output of a derivation chain. The learnable fusion weights w1, w2 and the classifier parameters are fit to training data in the standard supervised way, and the reported accuracies are subsequently evaluated on held-out configurations; no target quantity is defined in terms of a fitted parameter. The baselines are taken from Frank et al. (2023) and the DeepHS-Net papers, which are external works, not self-citations by the present authors, so there is no load-bearing self-citation. The adopted Frank et al. split is an external benchmark protocol, and while the paper does not describe the split or provide error bars, that is a reproducibility/statistical-significance concern rather than circularity. The inconsistency between Eq. (3) (a weighted sum f = w1*m + w2*s) and the text describing concatenation with an FC input dimension of [2λ, ...] is an internal ambiguity, but it does not make any result true by construction. No pattern of self-definition, fitted input called prediction, imported uniqueness, or renaming of a known result is present.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The method relies on standard signal processing and domain-specific assumptions about fruit spectra. The training hyperparameters are hand-chosen with no stated justification. No new physical entities are introduced.

free parameters (6)
  • learning_rate = 0.001
    Chosen by hand in Section 3.1, no justification given.
  • weight_decay = 1e-4
    Hand-chosen regularization term in Section 3.1.
  • batch_size = 16
    Hand-chosen in Section 3.1.
  • dropout_rate = 0.4
    Hand-chosen in Section 3.1.
  • epochs_per_config = not disclosed
    The paper states a variable number of epochs is 'determined experimentally' for each configuration, but no values are given.
  • fc_hidden_dimensions = [2λ, 512, 256]
    Architectural choice in Section 3.1, not derived from first principles.
assumptions (3)
  • domain assumption Fourier transform magnitude preserves ripeness-discriminative texture information
    Section 2.2 states frequency transformation makes periodic patterns of textures more discernible, but this is asserted without evidence.
  • domain assumption The central pixel of the fruit image is the most chemically informative point
    Section 2.2 asserts the central pixel is generally indicative of overall ripeness, but this is a heuristic with no validation.
  • domain assumption The DeepHS Fruit dataset labels and splits from Frank et al. are valid
    The paper relies on the dataset and the split protocol from Frank et al. without describing or verifying them.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fruit-HSNet: A Machine Learning Approach for Hyperspectral Image-Based Fruit Ripeness Prediction." pith.science (2026). https://pith.science/paper/JI6WN6T5

@misc{pith2026260801202,
  author       = {Pith},
  title        = {Pith review of: Fruit-HSNet: A Machine Learning Approach for Hyperspectral Image-Based Fruit Ripeness Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JI6WN6T5}},
  note         = {Machine review of arXiv:2608.01202}
}
read the original abstract

Fruit ripeness prediction (FRP) is a classification-based agricultural computer vision task that has attracted much attention, thanks to its wide-ranging advantages in agriculture field for both pre-harvest and post-harvest management. Accurate and timely FRP can be achieved using machine/deep learning-based hyperspectral image classification techniques. However, challenges including the limited availability of labeled data and the lack of robust methods generalizable to various hyperspectral cameras and fruit types can compromise the effectiveness of hyperspectral image-based FRP. Addressing these challenges, this paper introduces Fruit-HSNet, a machine learning architecture specifically designed for hyperspectral classification of fruit ripeness. Fruit-HSNet incorporates a spatio-spectral feature extraction module based on Fourier Transform and central pixel spectral signature followed by learnable feature fusion and a classifier optimized for ripeness classification. The proposed architecture was evaluated using the DeepHS Fruit dataset, the largest publicly available labeled real-world hyperspectral dataset for predicting fruit ripeness, which includes five different types of fruits-avocado, kiwi, mango, kaki, and papaya-captured with three distinct hyperspectral cameras at various stages of ripeness. Experimental results highlight that Fruit-HSNet substantially outperforms existing deep learning methods, from baseline to state-of-the-art models, with improvements of 12%, achieving a new state-of-the-art overall accuracy of 70.73%.

Figures

Figures reproduced from arXiv: 2608.01202 by the authors.

Figure 1
Figure 1. Illustration of Fruit-HSNet architecture. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Performance metrics of Fruit-HSNet across all [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Performance Evaluation of Fruit-HSNet for hyperspectral classification of fruit ripeness [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 18 canonical work pages

  1. [1]

    M., Mazzara, M., Li, C., Yao, J., Li, H., Aryal, J., Vivone, G., and Hong, D

    Ahmad, M., Distifano, S., Khan, A. M., Mazzara, M., Li, C., Yao, J., Li, H., Aryal, J., Vivone, G., and Hong, D. (2024). A comprehensive survey for hyperspectral image classification: The evolution from conventional to transformers.arXiv abs/2404.14955

    arXiv 2024

  2. [2]

    Chanussot, J. (2022). Hyperspectral image classifica- tion—traditional to deep models: A survey for future prospects.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 15:968–999

    work page 2022

  3. [3]

    and Trehan, U

    Chakraborty, T. and Trehan, U. (2021). Spectralnet: Ex- ploring spatial-spectral waveletcnn for hyperspectral image classification.ArXiv, abs/2104.00341

    arXiv 2021

  4. [4]

    A., and Zell, A

    Frank, H., Varga, L. A., and Zell, A. (2023). Hyperspec- tral benchmark: Bridging the gap between hsi applica- tions through comprehensive dataset and pretraining. arXiv preprint arXiv:2309.11122

    arXiv 2023

  5. [5]

    B., and Benediktsson, J

    Serpico, S. B., and Benediktsson, J. A. (2018). New frontiers in spectral-spatial hyperspectral image clas- sification: The latest advances based on mathemati- cal morphology, markov random fields, segmentation, sparse representation, and deep learning.IEEE Geo- science and Remote Sensing Magazine, 6(3):10–43

    work page 2018

  6. [6]

    Ghazal, S., Munir, A., and Qureshi, W. S. (2024). Com- puter vision in smart agriculture and precision farm- ing: Techniques and applications.Artificial Intelli- gence in Agriculture, 13:64–83

    work page 2024

  7. [7]

    He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid- ual learning for image recognition. In2016 IEEE Con- ference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV , USA, June 27-30, 2016, pages 770–778. IEEE Computer Society

    work page 2016

  8. [8]

    Hong, D., Han, Z., Yao, J., Gao, L., Zhang, B., Plaza, A., and Chanussot, J. (2022). Spectralformer: Rethinking hyperspectral image classification with transformers. IEEE Trans. Geosci. Remote. Sens., 60:1–15

    work page 2022

Show all 21 references
  1. [9]

    S., Rambabu, M., and Dua, Y

    Kumar, V ., Singh, R. S., Rambabu, M., and Dua, Y . (2024). Deep learning for hyperspectral image classification: A survey.Computer Science Review, 53:100658

    2024

  2. [10]

    Nalepa, J. (2020). Hyperspectral band selection using attention-based convolutional neural networks.IEEE Access, 8:42384–42403

    2020

  3. [11]

    and Young, S

    Lu, Y . and Young, S. (2020). A survey of public datasets for computer vision tasks in precision agriculture.Com- puters and Electronics in Agriculture, 178:105760

    2020

  4. [12]

    Luo, J., Li, B., and Leung, C. (2023). A survey of com- puter vision technologies in urban and controlled- environment agriculture.ACM Computing Surveys, 56(5):1–39

    2023

  5. [13]

    E., Haut, J

    Paoletti, M. E., Haut, J. M., Plaza, J., and Plaza, A. J. (2019). Deep learning classifiers for hyperspectral imaging: A review.Isprs Journal of Photogramme- try and Remote Sensing, 158:279–317. Pinto Barrera, J., Rueda-Chac ´on, H., Arguello, H., and

    2019

  6. [14]

    De, A. (2019). Classification of hass avocado (persea americana mill) in terms of its ripening via hyperspec- tral images.TecnoL ´ogicas, 22:109–128

    2019

  7. [15]

    Sun, X. (2024). A systematic review of hyperspectral imaging in precision agriculture: Analysis of its cur- rent state and future prospects.Computers and Elec- tronics in Agriculture, 222:109037

    2024

  8. [16]

    Albarelli, A. (2023). Fruit ripeness classification: A survey.Artificial Intelligence in Agriculture, 7:44–57

    2023

  9. [17]

    K., Krishna, G., Dubey, S

    Roy, S. K., Krishna, G., Dubey, S. R., and Chaudhuri, B. B. (2020). Hybridsn: Exploring 3-d-2-d CNN feature hi- erarchy for hyperspectral image classification.IEEE Geosci. Remote. Sens. Lett., 17(2):277–281

    2020

  10. [18]

    A., Frank, H., and Zell, A

    Varga, L. A., Frank, H., and Zell, A. (2023a). Self- supervised pretraining for hyperspectral classification of fruit ripeness. In6th International Conference on Optical Characterization of Materials, OCM 2023, pages 97–108

    2023

  11. [19]

    A., Makowski, J., and Zell, A

    Varga, L. A., Makowski, J., and Zell, A. (2021). Measuring the ripeness of fruit with hyperspectral imaging and deep learning. In2021 International Joint Conference on Neural Networks (IJCNN), pages 1–8

    2021

  12. [20]

    Yang, X., Cao, W., Lu, Y ., and Zhou, Y . (2022). Hyperspec- tral image transformer classification networks.IEEE Trans. Geosci. Remote. Sens., 60:1–15

    2022

  13. [21]

    Zhu, H., Chu, B., Fan, Y ., Tao, X., Yin, W., and He, Y . (2017). Hyperspectral imaging for predicting the in- ternal quality of kiwifruits based on variable selection algorithms and chemometric models.Scientific Re- ports, 7

    2017

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.