pith. sign in

arxiv: 2606.26757 · v1 · pith:PMG2H2XBnew · submitted 2026-06-25 · 💻 cs.LG

Batch-Invariant Spectral Intelligence for Robust and Explainable Insect Authentication

Majharulislam Babor , Giacomo Rossi , Annalisa Altavilla , Oliver Schl\"uter , Marina M.-C. H\"ohne This is my paper

Pith reviewed 2026-06-26 05:09 UTC · model grok-4.3

classification 💻 cs.LG
keywords near-infrared spectroscopyinsect species authenticationbatch-invariant learningadversarial trainingexplainable AIfood authenticationspectral preprocessing
0
0 comments X

The pith

The Batch-Invariant Spectral Network suppresses batch-specific variation in near-infrared spectra before learning species features, reaching 0.93 accuracy on unseen production batches.

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

The paper introduces BISN to authenticate edible insect species from near-infrared spectra that must work reliably across different production batches. It pairs a learnable preprocessing module, started from Savitzky-Golay filtering, with an entropy-regularized adversarial objective that removes batch signals early. This ordering lets the model focus on species traits rather than measurement artifacts. On 2700 spectra from three insect species across three batches, the method reaches 0.93 mean leave-one-batch-out accuracy while explanations point to lipid and protein regions. A reader would care because such robustness supports safe use of insects as food without retraining for every new batch.

Core claim

BISN is an end-to-end framework that combines a learnable preprocessing module, initialised with Savitzky-Golay filtering, with an entropy-regularised adversarial objective to suppress batch-specific spectral variation before species-specific features are learned. Using 2,700 spectra from three species collected across three independent production batches, BISN achieves a mean leave-one-batch-out accuracy of 0.93 (standard deviation 0.04), outperforming the strongest baseline by four percent. Explainable AI shows that model decisions consistently rely on the lipid and protein absorption regions across all folds.

What carries the argument

The Batch-Invariant Spectral Network (BISN), an end-to-end architecture that applies entropy-regularized adversarial suppression of batch effects prior to species feature extraction on a learnable preprocessing module.

If this is right

  • BISN outperforms domain-adversarial networks that apply adaptation only after feature extraction.
  • Accuracy remains high on completely unseen production batches with low variance across folds.
  • Explanations tie predictions to lipid and protein absorption regions in every cross-batch test.
  • The method enables automated species authentication under realistic industrial batch variation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The early-suppression design may transfer to other spectroscopic classification tasks that suffer from instrument or batch drift.
  • Public release of code and data allows direct testing on new insect species or spectrometer models.
  • If the method preserves species information while removing batch signals, similar preprocessing could shorten the data collection needed for new food-authentication problems.

Load-bearing premise

The entropy-regularized adversarial objective removes batch-specific spectral variation without discarding the information required for accurate species discrimination.

What would settle it

Retraining the model with the adversarial term removed or with its strength varied and observing whether leave-one-batch-out accuracy falls to or below the strongest baseline level.

Figures

Figures reproduced from arXiv: 2606.26757 by Annalisa Altavilla, Giacomo Rossi, Majharulislam Babor, Marina M.-C. H\"ohne, Oliver Schl\"uter.

Figure 1
Figure 1. Figure 1: Architecture of the Batch-Invariant Spectral Network (BISN). Raw NIR spectra 𝑥 enter an informed preprocessing module consisting of a Savitzky–Golay-initialised learnable 1D convolution followed by instance normalisation, producing a batch-invariant representation ̂𝐱. This representation feeds into a sparse attentive encoder that generates a compact latent embedding 𝐳 for insect species classification. Spe… view at source ↗
Figure 2
Figure 2. Figure 2: Spectral sensitivity analysis across insect species, processing treatments, and chemically defined NIR wavelength regions. (a) Mean raw NIR spectra showing species-specific baseline profiles and wavelength-dependent variability. (b) The shaded bands indicate the eleven chemically defined regions detailed in Supplementary S4. (c–e) Region-level spectral perturbation heatmaps for insect species. The rows cor… view at source ↗
Figure 3
Figure 3. Figure 3: NIR spectral profiles across batches, species, and preprocessing stages. Top row: mean spectra coloured by species identity (A. domesticus: blue, H. illucens: orange, T. molitor: purple) for (a) raw, (b) classically preprocessed, (c) BISN batch-invariant representations, and (d) Explained variance ratio for the first ten principal components across the three representations, showing that classical preproce… view at source ↗
Figure 4
Figure 4. Figure 4: BISN latent embedding visualisation across leave-one-batch-out (LOBO) folds. (a–c) PCA projections of the 8-dimensional BISN embedding for LOBO folds. Marker shapes denote batch identity and colours denote species. (d–f) Cosine similarity matrices of mean BISN embeddings across insect species for LOBO folds 1, 2, and 3. AD: A. domesticus; HI: H. illucens; TM: T. molitor. Off-diagonal values reflect inter-s… view at source ↗
Figure 5
Figure 5. Figure 5: d. The H. illucens→T. molitor axis accounts for 54.7 % of all errors, driven by blanching-induced thermal denaturation and moisture redistribution, which reduce the compositional contrast between these two species. Under blanching with ultrasound (T1_U1), T. molitor → H. illucens errors rise from 7.3 % to 16.9 % while H. illucens → T. molitor decreases from 11.3 % to 7.3 %, an asymmetry consistent with cav… view at source ↗
Figure 6
Figure 6. Figure 6: Spectral attribution and counterfactual analysis of BISN predictions. (a–c) Mean Integrated Gradients (IG) attribution per spectral region for each species. (d) IG attribution distributions across all LOBO folds. (e) Mean perturbation magnitude per spectral region against true-class recovery rate, defined as the proportion of misclassified samples across all three LOBO folds for which optimising only that … view at source ↗
read the original abstract

Edible insects offer an efficient source of alternative protein, requiring less land, water and emitting less greenhouse gas than conventional livestock. However, their successful integration into the food supply chain demands reliable species authentication to control allergen exposure, prevent adulteration, and meet regulatory standards. Near-infrared spectroscopy provides a rapid analytical tool, but its performance drops when applied to production batches unseen during training due to batch-to-batch variation in spectral measurements. We introduce the Batch-Invariant Spectral Network (BISN), an end-to-end framework that combines a learnable preprocessing module, initialised with Savitzky-Golay filtering, with an entropy-regularised adversarial objective to suppress batch-specific spectral variation. In contrast to Domain-Adversarial Neural Networks, which enforce domain adaptation only after feature extraction, BISN suppress batch-effects before species-specific features are learned. Using 2,700 spectra from three species (Acheta domesticus, Hermetia illucens, and Tenebrio molitor) collected across three independent production batches, BISN achieves a mean leave-one-batch-out accuracy of 0.93 (standard deviation 0.04), outperforming the strongest baseline by four percent. Further insights gained by using explainable AI confirm that model decisions consistently rely on the lipid and protein absorption regions across all folds, connecting predictive performance to known insect biochemistry. BISN addresses both cross-batch robustness and biochemical interpretability for automated insect species authentication under realistic industrial conditions. The source code and dataset are publicly available at https://github.com/majharB/bisn.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces the Batch-Invariant Spectral Network (BISN), an end-to-end architecture that pairs a learnable Savitzky-Golay preprocessing module with an entropy-regularized adversarial objective to suppress batch effects in NIR spectra prior to species feature extraction. On 2,700 spectra from three insect species collected over three production batches, BISN reports a mean leave-one-batch-out accuracy of 0.93 (sd 0.04), a 4% improvement over the strongest baseline, with XAI attributions consistently highlighting lipid and protein absorption bands.

Significance. If the adversarial regularization demonstrably preserves species-discriminative information while removing batch variation, the work supplies a concrete, interpretable pipeline for industrial insect authentication that links model decisions to known biochemistry. Public code and data release is a clear strength that supports reproducibility and extension.

major comments (2)
  1. [Abstract / Method] Abstract and Method section: The central claim that the entropy-regularized adversarial objective removes batch-specific spectral variation before species features are learned rests on an unverified assumption. No quantitative check (post-preprocessing batch classification accuracy, mutual information with batch labels, or ablation over regularization strength) is reported despite only three batches being available; this directly affects whether the 0.93 LOO accuracy can be attributed to the proposed mechanism rather than the preprocessing module alone.
  2. [Results] Results section: With only three batches and three species, the leave-one-batch-out protocol provides limited statistical power; the reported standard deviation of 0.04 is given but no per-fold confusion matrices, per-batch accuracies, or statistical significance tests against baselines are described, weakening the robustness claim.
minor comments (2)
  1. [Abstract] Abstract: The number of spectra per species and per batch is not stated, making it difficult to assess class balance and batch-size effects.
  2. [Related Work / Experiments] The distinction from Domain-Adversarial Neural Networks is conceptually clear but would benefit from an explicit side-by-side result table showing whether pre-feature suppression yields gains beyond standard DANN.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on verifying the batch-invariance mechanism and the statistical details of the evaluation. We address each major comment below and outline planned revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract / Method] Abstract and Method section: The central claim that the entropy-regularized adversarial objective removes batch-specific spectral variation before species features are learned rests on an unverified assumption. No quantitative check (post-preprocessing batch classification accuracy, mutual information with batch labels, or ablation over regularization strength) is reported despite only three batches being available; this directly affects whether the 0.93 LOO accuracy can be attributed to the proposed mechanism rather than the preprocessing module alone.

    Authors: We agree that the manuscript would benefit from explicit quantitative checks to support attribution of the performance gains to the adversarial objective. In the revised manuscript we will add post-preprocessing batch classification accuracy, estimates of mutual information between the learned features and batch labels, and an ablation over regularization strength. These additions will demonstrate that the entropy-regularized term contributes to batch-effect suppression beyond the learnable Savitzky-Golay preprocessing module alone. revision: yes

  2. Referee: [Results] Results section: With only three batches and three species, the leave-one-batch-out protocol provides limited statistical power; the reported standard deviation of 0.04 is given but no per-fold confusion matrices, per-batch accuracies, or statistical significance tests against baselines are described, weakening the robustness claim.

    Authors: We acknowledge that the small number of batches inherently limits statistical power. The revised manuscript will include per-fold confusion matrices, per-batch accuracies for each leave-one-batch-out fold, and statistical significance tests (e.g., McNemar’s test) comparing BISN against the baselines. These additions will provide greater transparency on variability across folds while noting the dataset constraints. revision: yes

Circularity Check

0 steps flagged

No significant circularity: empirical evaluation on external data with independent architecture

full rationale

The paper introduces BISN as an end-to-end architecture with learnable Savitzky-Golay preprocessing and entropy-regularized adversarial training, then reports an empirical leave-one-batch-out accuracy of 0.93 on a held-out dataset of 2700 spectra from three batches. This accuracy is obtained via standard cross-validation on external measurements rather than being defined by or reduced to the method's own parameters or equations. No self-citations, uniqueness theorems, or fitted-input-as-prediction patterns appear in the abstract or described claims. The XAI consistency with lipid/protein regions is an interpretive post-hoc analysis, not a load-bearing derivation step. The framework is self-contained against public data and code without reducing the central performance claim to a tautology.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the premise that adversarial entropy regularization can isolate batch effects early without harming species discrimination; this is a domain assumption rather than a derived result.

axioms (1)
  • domain assumption Adversarial training with entropy regularization can remove batch-specific information from spectra prior to species classification without degrading discriminative power for the target task.
    Core design choice of the BISN framework stated in the abstract.

pith-pipeline@v0.9.1-grok · 5828 in / 1351 out tokens · 28021 ms · 2026-06-26T05:09:13.619426+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 35 canonical work pages

  1. [1]

    M.Psarianos,F.Aghababaei,O.K.Schlüter,Bioactivecompoundsinedibleinsects: aspectsofcultivation,processingandnutrition, FoodResearchInternational,203(2025)115802.https://doi.org/10.1016/j.foodres.2025.115802

    work page doi:10.1016/j.foodres.2025.115802 2025

  2. [2]

    Lange, Y

    K.W. Lange, Y. Nakamura, Edible insects as future food: chances and challenges,Journal of Future Foods, 1 (2021) 38–46. https://doi.org/10.1016/j.jfutfo.2021.10.001

    work page doi:10.1016/j.jfutfo.2021.10.001 2021

  3. [3]

    G.Rossi,S.Ojha,J.Hankel,O.K.Schlüter,Insect-mediatedvalorisationofanaerobicallydigestedaquaculturewaste: bioconversion performances,nutritionalcompositionandmicrobialsafetyofblacksoldierflylarvae,SustainableFoodTechnology,3(2025) 811–821.https://doi.org/10.1039/D4FB00392F

    work page doi:10.1039/d4fb00392f 2025

  4. [4]

    F.A.Madau,B.Arru,R.Furesi,P.Pulina,F.Riu,Insectfarmingforfeedandfoodproductionfromacircularbusinessmodel perspective,Sustainability,12(2020)5418.https://doi.org/10.3390/su12135418

    work page doi:10.3390/su12135418 2020

  5. [5]

    https://doi.org/10.3390/foods14050751

    M.Wildbacher,J.Andronache,K.Pühringer,S.Dobrovolny,R.Hochegger,M.Cichna-Markl,AuthenticationofEU-authorized edible insect species in food products by DNA barcoding and high-resolution melting (HRM) analysis,Foods, 14 (2025) 751. https://doi.org/10.3390/foods14050751

    work page doi:10.3390/foods14050751 2025

  6. [6]

    Lawal, R.R

    K.G. Lawal, R.R. Kavle, T.O. Akanbi, M. Mirosa, D. Agyei, Lipid nutritional indices, regioisomeric distribution, and thermal propertiesofTenebriomolitorandHermetiaillucenslarvaefat,JournalofAsia-PacificEntomology,25(2022)101951. https: //doi.org/10.1016/j.aspen.2022.101951

    work page doi:10.1016/j.aspen.2022.101951 2022

  7. [7]

    3389/fmicb.2021.620798

    L.Borrelli,L.Varriale,L.Dipineto,A.Pace,L.F.Menna,A.Fioretti,Insectderivedlauricacidaspromisingalternativestrategy to antibiotics in the antimicrobial resistance scenario,FrontiersinMicrobiology, 12 (2021) 620798.https://doi.org/10. 3389/fmicb.2021.620798

    arXiv 2021

  8. [8]

    Laroche, V

    M. Laroche, V. Perreault, A. Marciniak, A. Gravel, J. Chamberland, A. Doyen, Comparison of conventional and sustainable lipid extraction methods for the production of oil and protein isolate from edible insect meal,Foods, 8 (2019) 572.https: //doi.org/10.3390/foods8110572

    work page doi:10.3390/foods8110572 2019

  9. [9]

    ISBN978-92-5-107596-8.https://www.fao.org/4/i3253e/i3253e.pdf

    A.vanHuis,J.VanItterbeeck,H.Klunder,E.Mertens,A.Halloran,G.Muir,P.Vantomme,EdibleInsects: FutureProspectsfor FoodandFeedSecurity,FAOForestryPaper,vol.171,FoodandAgricultureOrganizationoftheUnitedNations,Rome,2013. ISBN978-92-5-107596-8.https://www.fao.org/4/i3253e/i3253e.pdf

    2013

  10. [10]

    B.A.Rumpold,O.K.Schlüter,Nutritionalcompositionandsafetyaspectsofedibleinsects,MolecularNutrition&FoodResearch, 57(2013)802–823.https://doi.org/10.1002/mnfr.201200735

    work page doi:10.1002/mnfr.201200735 2013

  11. [12]

    Chakrabarty, C.K

    S. Chakrabarty, C.K. Deb, S. Marwaha, M.A. Haque, D. Kamil, R. Bheemanahalli, P.R. Shashank, Application of artificial intelligenceininsectpestidentification–areview,ArtificialIntelligenceinAgriculture,16(2026)44–61. https://doi.org/ 10.1016/j.aiia.2025.06.005

    work page doi:10.1016/j.aiia.2025.06.005 2026

  12. [13]

    J.H. Qu, D. Liu, J.H. Cheng, D.W. Sun, J. Ma, H. Pu, X.A. Zeng, Applications of near-infrared spectroscopy in food safety evaluationandcontrol: areviewofrecentresearchadvances,CriticalReviewsinFoodScienceandNutrition,60(2020)2764–2796. https://doi.org/10.1080/10408398.2019.1651248

    work page doi:10.1080/10408398.2019.1651248 2020

  13. [14]

    https://doi.org/10.1016/j.tifs.2021.11

    L.Zhang,X.Huang,Z.Li,J.Zhao,Q.Zhang,Areviewofnear-infraredspectroscopyforfoodauthentication: challengesand perspectives,TrendsinFoodScience&Technology,118(2021)806–820. https://doi.org/10.1016/j.tifs.2021.11. 010

    work page doi:10.1016/j.tifs.2021.11 2021

  14. [15]

    D.A.Burns,E.W.Ciurczak(Eds.),HandbookofNear-InfraredAnalysis,3rded.,CRCPress,BocaRaton,FL,USA,2007

    2007

  15. [17]

    Y.Xu,T.Kong,Y.Ma,Y.Zhao,L.Chu,M.Zheng,Near-infraredspectroscopy: applicationinensuringfoodqualityandsafety, AnalyticalMethods,17(2025)3381–3406.https://doi.org/10.1039/D4AY02039A

    work page doi:10.1039/d4ay02039a 2025

  16. [18]

    https://doi.org/10.1016/j.saa.2024.125646

    J.P.Cruz-Tirado,M.S.dosSantosVieira,R.S.B.Ferreira,J.M.Amigo,E.A.C.Batista,D.F.Barbin,Predictionoftotallipidsandfatty acidsinblacksoldierfly(HermetiaillucensL.)driedlarvaebyNIR-hyperspectralimagingandchemometrics,SpectrochimicaActa PartA:MolecularandBiomolecularSpectroscopy,329(2025)125646. https://doi.org/10.1016/j.saa.2024.125646

    work page doi:10.1016/j.saa.2024.125646 2025

  17. [19]

    S.Tan,S.Hu,S.He,L.Zhu,Y.Qian,Y.Deng,Leveraginghyperspectralimagesforaccurateinsectclassificationwithanovel two-branchself-correlationapproach,Agronomy,14(2024)863.https://doi.org/10.3390/agronomy14040863

    work page doi:10.3390/agronomy14040863 2024

  18. [20]

    Cruz-Tirado, J.M

    J.P. Cruz-Tirado, J.M. Amigo, D.F. Barbin, Determination of protein content in single black soldier fly (Hermetia illucens L.) larvae by near infrared hyperspectral imaging (NIR-HSI) and chemometrics,Food Control, 143 (2023) 109266.https: //doi.org/10.1016/j.foodcont.2022.109266

    work page doi:10.1016/j.foodcont.2022.109266 2023

  19. [21]

    Rinnan, F

    Å. Rinnan, F. van den Berg, S.B. Engelsen, Review of the most common pre-processing techniques for near-infrared spectra, TrACTrendsinAnalyticalChemistry,28(2009)1201–1222.https://doi.org/10.1016/j.trac.2009.07.007

    work page doi:10.1016/j.trac.2009.07.007 2009

  20. [22]

    Y.Zhao,J.Yu,P.Shan,Z.Zhao,X.Jiang,S.Gao,PLSsubspace-basedcalibrationtransferfornear-infraredspectroscopyquantitative analysis,Molecules,24(2019)1289.https://doi.org/10.3390/molecules24071289

    work page doi:10.3390/molecules24071289 2019

  21. [23]

    D.Azzollini,A.Derossi,V.Fogliano,C.M.M.Lakemond,C.Severini,Effectsofformulationandprocessconditionsonmicrostruc- ture, texture and digestibility of extruded insect-riched snacks,Innovative Food Science & Emerging Technologies, 45 (2018) 344–353.https://doi.org/10.1016/j.ifset.2017.11.017

    work page doi:10.1016/j.ifset.2017.11.017 2018

  22. [24]

    Ni, S.-Y

    J.-B. Ni, S.-Y. Luo, Y.-X. Bi, S. Zielinska, C.-J. Ding, J.-L. Tao, Z. Ning, W.-L. Tian, W.-J. Peng, X.-M. Fang, The combined effectsofultrasoundandplasma-activatedwateronsilkwormpupae: physicochemicalproperties,microbiologicaldiversityand ultrastructure,UltrasonicsSonochemistry,107(2024)106927. https://doi.org/10.1016/j.ultsonch.2024.106927

    work page doi:10.1016/j.ultsonch.2024.106927 2024

  23. [25]

    https://doi.org/10.1021/ac00023a016

    Y.Wang, D.J.Veltkamp, B.R.Kowalski, Multivariateinstrumentstandardization,AnalyticalChemistry, 63(1991)2750–2756. https://doi.org/10.1021/ac00023a016

    work page doi:10.1021/ac00023a016 1991

  24. [26]

    R.Nikzad-Langerodi,W.Zellinger,E.Lughofer,S.Saminger-Platz,Domain-invariantpartial-least-squaresregression,Analytical Chemistry,90(2018)6693–6701.https://doi.org/10.1021/acs.analchem.8b00498

    work page doi:10.1021/acs.analchem.8b00498 2018

  25. [27]

    1016/j.rineng.2025.108477

    M.Babor,S.Liu,A.Arefi,A.Olszewska-Widdrat,J.Venus,B.Sturm,M.M.-C.Höhne,Interpretabledomainadaptationenables robust lactic acid fermentation monitoring from waste,Results in Engineering, 29 (2026) 108477.https://doi.org/10. 1016/j.rineng.2025.108477

    arXiv 2026

  26. [28]

    S.Ö.Arik,T.Pfister,TabNet: attentiveinterpretabletabularlearning,in:Proceedingsofthe35thAAAIConferenceonArtificial Intelligence,35(2021)6679–6687.https://doi.org/10.1609/aaai.v35i8.16826

    work page doi:10.1609/aaai.v35i8.16826 2021

  27. [29]

    http://jmlr.org/papers/v17/15-239

    Y.Ganin,E.Ustunova,H.Ajakan,P.Germain,H.Larochelle,F.Laviolette,M.Marchand,V.Lempitsky,Domain-adversarialtrain- ingofneuralnetworks,JournalofMachineLearningResearch,17(2016)1–35. http://jmlr.org/papers/v17/15-239. html

    2016

  28. [30]

    https: //doi.org/10.1016/j.ijbiomac.2024.135057

    X.Ni,C.Chen,R.Li,Q.Liu,C.Duan,X.Wang,M.Xu,Effectsofultrasonictreatmentonthestructureandfunctionalcharacteristics ofmyofibrillarproteinsfromblacksoldierfly,InternationalJournalofBiologicalMacromolecules,278(2024)135057. https: //doi.org/10.1016/j.ijbiomac.2024.135057

    work page doi:10.1016/j.ijbiomac.2024.135057 2024

  29. [31]

    https://doi.org/10

    A.M.Bashkatov,E.A.Genina,V.I.Kochubey,V.V.Tuchin,Opticalpropertiesofhumanskin,subcutaneousandmucoustissuesin thewavelengthrangefrom400to2000nm,JournalofPhysicsD:AppliedPhysics,38(2005)2543–2555. https://doi.org/10. 1088/0022-3727/38/15/004

    2005

  30. [32]

    https://doi.org/10.1255/jnirs.869

    R.Tsenkova,Aquaphotomics: waterabsorbancepatternasabiologicalmarker,NearInfraredSpectroscopy,17(2009)303–314. https://doi.org/10.1255/jnirs.869

    work page doi:10.1255/jnirs.869 2009

  31. [33]

    R.J.Barnes,M.S.Dhanoa,S.J.Lister,Standardnormalvariatetransformationandde-trendingofnear-infrareddiffusereflectance spectra,AppliedSpectroscopy,43(1989)772–777.https://doi.org/10.1366/0003702894202201

    work page doi:10.1366/0003702894202201 1989

  32. [34]

    A.Savitzky,M.J.E.Golay,Smoothinganddifferentiationofdatabysimplifiedleastsquaresprocedures,AnalyticalChemistry,36 (1964)1627–1639.https://doi.org/10.1021/ac60214a047. 19

    work page doi:10.1021/ac60214a047 1964

  33. [35]

    A.Gretton,K.M.Borgwardt,M.J.Rasch,B.Schölkopf,A.Smola,Akerneltwo-sampletest,JournalofMachineLearningResearch, 13(2012)723–773.https://jmlr.org/papers/v13/gretton12a.html

    2012

  34. [36]

    P.J.Rousseeuw,Silhouettes: agraphicalaidtotheinterpretationandvalidationofclusteranalysis,JournalofComputationaland AppliedMathematics,20(1987)53–65.https://doi.org/10.1016/0377-0427(87)90125-7

    work page doi:10.1016/0377-0427(87)90125-7 1987

  35. [37]

    Martins, R.F

    A.F.T. Martins, R.F. Astudillo, From softmax to sparsemax: a sparse model of attention and multi-label classification, in: Proceedingsofthe33rdInternationalConferenceonMachineLearning,PMLR48(2016)1614–1623. https://proceedings. mlr.press/v48/martins16.html

    2016

  36. [38]

    Incomplete Randomized Blocks

    R.A.Fisher,Theuseofmultiplemeasurementsintaxonomicproblems,AnnalsofEugenics,7(1936)179–188. https://doi. org/10.1111/j.1469-1809.1936.tb02137.x

    work page doi:10.1111/j.1469-1809.1936.tb02137.x 1936

  37. [39]

    https: //gaussianprocess.org/gpml

    C.E.Rasmussen, C.K.I.Williams,GaussianProcessesforMachineLearning, MITPress, Cambridge, MA,USA,2006. https: //gaussianprocess.org/gpml

    2006

  38. [40]

    Barker, W

    M. Barker, W. Rayens, Partial least squares for discrimination,Journal of Chemometrics, 17 (2003) 166–173.https://doi. org/10.1002/cem.785

    work page doi:10.1002/cem.785 2003

  39. [41]

    P. Fu, Y. Wen, Y. Zhang, L. Li, Y. Feng, L. Yin, H. Yang, SpectraTr: a novel deep learning model for qualitative analysis of drug spectroscopy based on transformer structure,Journal of Innovative Optical Health Sciences, 15 (2022) 2250021.https: //doi.org/10.1142/S1793545822500213

    work page doi:10.1142/s1793545822500213 2022

  40. [42]

    N.Singh,S.Kaur,S.R.Padhi,R.John,A.Kumar,J.C.Rana,R.Bhardwaj,A.Riar,NIRCoreVision: anoveldeeplearning-based framework with GUI integration for core set selection from NIRS data using 1D CNN and k-means clustering,Journal of AgricultureandFoodResearch,24(2025)102390.https://doi.org/10.1016/j.jafr.2025.102390

    work page doi:10.1016/j.jafr.2025.102390 2025

  41. [43]

    Hollmann, S

    N. Hollmann, S. Müller, L. Purucker, A. Krishnakumar, M. Körfer, S.B. Hoo, R.T. Schirrmeister, F. Hutter, Accurate pre- dictions on small data with a tabular foundation model,Nature, 637 (2025) 1043–1049.https://doi.org/10.1038/ s41586-024-08328-6

    2025

  42. [44]

    https://doi.org/10.1016/j.foodcont.2023.110268

    D.Ni,J.L.D.Nelis,A.L.Dawson,N.Bourne,P.Juliano,M.L.Colgrave,A.Juhász,U.Bose,Applicationofnear-infraredspectroscopy andchemometricsfortherapiddetectionofinsectproteinadulterationfromasimulatedmatrix,FoodControl,159(2024)110268. https://doi.org/10.1016/j.foodcont.2023.110268

    work page doi:10.1016/j.foodcont.2023.110268 2024

  43. [45]

    X.Zhang,F.Yang,J.Xiao,H.Qu,N.F.Jocelin,L.Ren,Y.Guo,Analysisandcomparisonofmachinelearningmethodsforspecies identificationutilizingATR-FTIRspectroscopy,SpectrochimicaActaPartA:MolecularandBiomolecularSpectroscopy,308(2024) 123713.https://doi.org/10.1016/j.saa.2023.123713

    work page doi:10.1016/j.saa.2023.123713 2024

  44. [46]

    Sampling-Based Risk-Aware Path Planning Around Dynamic Engagement Zones,

    Z.Ma,M.Di,T.Hu,X.Wang,J.Zhang,Z.He,Visible-NIRhyperspectralimagingbasedoncharacteristicspectraldistillationused forspeciesidentificationofsimilarcrickets,Optics&LaserTechnology,183(2025)112420. https://doi.org/10.1016/j. optlastec.2025.112420

    work page doi:10.1016/j 2025

  45. [47]

    https://proceedings.mlr.press/v70/sundararajan17a

    M.Sundararajan,A.Taly,Q.Yan,Axiomaticattributionfordeepnetworks,in:Proceedingsofthe34thInternationalConferenceon MachineLearning(ICML),PMLR70(2017)3319–3328. https://proceedings.mlr.press/v70/sundararajan17a. html

    2017

  46. [48]

    https://doi.org/10.1007/s11947-024-03510-0

    N.Muñoz-Seijas,H.Fernandes,J.M.Domínguez,J.M.Salgado,RecentadvancesinbiorefineryofTenebriomolitoradoptinggreen technologies,FoodandBioprocessTechnology18,1061–1078(2025). https://doi.org/10.1007/s11947-024-03510-0

    work page doi:10.1007/s11947-024-03510-0 2025

  47. [49]

    Samples = 50

    E. Hoffer, R. Banner, I. Golan, D. Soudry, Train longer, generalise better: closing the generalisation gap in large batch train- ing of neural networks, in:Advances in Neural Information Processing Systems (NeurIPS), 30 (2017) 1731–1741.https: //proceedings.neurips.cc/paper/2017/file/a5e0ff62be0b08456fc7f1e88812af3d-Paper.pdf. 20 SupplementaryMaterials S1...

    2017