Non-destructive identification of commercial jerky types based on multi-band hyperspectral imaging with machine learning

Yuanxi Han, Liang Li, Siyuan Jiang, Pengpeng Sun, Wenliang Wu, Zhendong LiuFood 2025 · DOI

TL;DR

This study leverages multi-band hyperspectral imaging (Vis-NIR and SWIR) combined with machine learning to non-destructively identify five common types of commercial jerky. The research highlights the effectiveness of combining normalization and randomization preprocessing, and identifies Logistic Regression and linear-kernel SVM as top-performing models, especially when using SWIR data, offering a promising solution to combat jerky fraud.

What Problem Does This Paper Address?

The commercial jerky market faces a significant problem with counterfeiting and fraudulent labeling. Consumers often struggle to distinguish between different types of jerky (e.g., beef vs. yak, or duck vs. pork) due to the convergence of flavor and appearance after processing. This discrepancy in market value incentivizes unscrupulous merchants to misrepresent products, leading to economic losses for consumers and legitimate businesses. Existing methods for authenticity testing, such as traditional physicochemical or nucleic acid analyses, often require destructive sample preparation and can be time-consuming. This paper aims to address this by developing a rapid, non-destructive method for identifying commercial jerky types using hyperspectral imaging and machine learning.

Methods

The core of this research lies in the application of hyperspectral imaging (HSI) and machine learning (ML) for classification.

  • Hyperspectral Imaging (HSI): The study employed two HSI systems:
  • Visible-Near-Infrared (Vis-NIR): Covering the 400-1000 nm range, this system captures spectral information related to the color of the jerky.
  • Short-Wave Infrared (SWIR): Covering the 900-1700 nm range, this system captures spectral information related to molecular composition, such as moisture and fat content.
  • Data Acquisition: Samples were arranged and imaged under controlled conditions, followed by calibration using dark and white reference images to minimize noise.
  • Machine Learning (ML) for Classification: Six common ML algorithms were evaluated:
  • Support Vector Machine (SVM)
    • Artificial Neural Network (ANN)
    • Naive Bayes (NB)
    • Logistic Regression (LR)
    • Decision Tree (DT)
    • Random Forest (RF)
  • Preprocessing Techniques: To enhance data quality and model performance, several preprocessing steps were investigated:
  • Normalization: Standardizing spectral intensity ranges to reduce instrumental variations and batch effects.
  • Randomization: Randomly partitioning data to minimize order dependence and improve generalization.
  • Combinations: The study compared scenarios with no preprocessing, normalization only, and normalization with randomization.
  • Model Optimization: Hyperparameter tuning was performed for key algorithms (SVM, ANN, LR, DT, RF) to find optimal settings for improved classification accuracy. This involved exploring different kernel functions for SVM, activation functions for ANN, solvers for LR, and split criteria for DT/RF.
  • Evaluation Metrics: Model performance was assessed using accuracy, precision, recall, and F1-score, with 10x10-fold cross-validation employed for robust evaluation.

Datasets

The study utilized a dataset of 150 commercial jerky samples, comprising 30 samples each of:

  • Chicken jerky
  • Yak jerky
  • Beef jerky
  • Duck jerky
  • Pork jerky

These samples were uniformly processed by Tibet Tianhu Herdsman Agriculture and Animal Husbandry Science and Technology Development Company Limited, ensuring consistency. The dataset was divided into training (90%) and testing (10%) sets with a fixed random seed for reproducibility. No new benchmark datasets were introduced; rather, existing commercial jerky products were used for evaluation.

Key Insights and Ideas

  • Preprocessing is Crucial: Both normalization and randomization preprocessing significantly improved model accuracy and stability across most algorithms and spectral bands. The combination of both proved most effective.
  • SWIR Band's Advantage: The SWIR band, capturing molecular composition, showed a particular advantage in differentiating certain jerky types (e.g., duck jerky) compared to the Vis-NIR band, which is more sensitive to color.
  • Linear Models Shine: Despite the complexity of hyperspectral data, linear models like Logistic Regression and SVM with a linear kernel demonstrated superior performance, especially after optimization, suggesting that the spectral features relevant for jerky identification might be linearly separable or can be effectively handled by linear transformations.
  • Complementary Spectral Information: The Vis-NIR and SWIR bands offer complementary information. While Vis-NIR excels at identifying chicken jerky (likely due to color differences), SWIR provides better differentiation for other types, indicating that fusing data from both bands could further enhance identification.
  • Hyperparameter Optimization Matters: Tuning hyperparameters, particularly for non-linear models like SVM, can significantly boost performance. Using a linear kernel for SVM, for instance, dramatically improved its accuracy.

Key Results

  • Overall Accuracy: After preprocessing and optimization, the Logistic Regression model achieved the highest overall identification accuracy, ranging from 85.78% to 100.00%.
  • SVM Performance: With hyperparameter optimization (linear kernel), SVM achieved high accuracies of 89.29% and 95.29% across the two spectral bands.
  • ANN Performance: The ANN with a hyperbolic tangent activation function achieved training performance exceeding 90.00% accuracy.
  • Preprocessing Impact: Normalization and randomization preprocessing consistently improved test set accuracies for most algorithms. For example, the average accuracy for LR on the test set increased from 65.00% (no preprocessing) to 91.92% (with normalization and randomization).
  • Band-Specific Strengths:
    • Vis-NIR band showed high accuracy for chicken jerky (often >90%).
  • SWIR band showed better differentiation for duck jerky and generally improved prediction accuracy for several types compared to Vis-NIR.
  • Algorithm Performance: LR and linear-kernel SVM consistently performed well across multiple jerky types, especially with SWIR data. NB showed the lowest average accuracy for most types. DT had high training accuracy but poor test accuracy, indicating overfitting.

Limitations and Future Work

  • Limitations:
    • The study focused on five specific types of commercially available jerky. The generalizability to other jerky types or homemade jerky was not assessed.
  • While the study explored different preprocessing and optimization strategies, the impact of other advanced techniques (e.g., deep learning architectures beyond basic ANN) was not investigated.
  • The paper notes that while DT and RF achieved 100% training accuracy, their prediction performance was lower, suggesting potential overfitting or sensitivity to noise.
  • Future Work:
    • The authors suggest exploring the effective fusion of spectral data and images from both Vis-NIR and SWIR bands to further enhance the precise identification of jerky varieties.
  • Investigating more advanced machine learning models, potentially including deep learning architectures, could lead to even higher identification accuracies.
  • Expanding the study to include a wider range of jerky types and processing methods would improve the robustness and applicability of the developed method.

Summary updated

Abstract

Commercial jerky counterfeiting is widespread in the market. This study combined visible-near-infrared and short-wave-near-infrared hyperspectral imaging along with multiple machine learning algorithms for non-destructive identification of five types of commercial jerky products, and explored the impact of different spectral bands, algorithm selection, and optimization methods on identification performance. After data preprocessing, all models' accuracies and stability improved. Specifically, the logistic regression model was best for jerky identification, with 85.78 %-100.00 % accuracy. With hyperparameter optimization, Support Vector Machine with linear kernel had highest accuracy (89.29 % and 95.29 % in two bands). Additionally, the artificial neural network with the hyperbolic tangent activation function had optimal training performance, exceeding 90.00 % accuracy. The findings demonstrate short-wave-near-infrared hyperspectral imaging combined with linear models (logistic regression and Support Vector Machine with linear kernel parameter settings) is better for identifying the types of jerky. This study provides reference for the band, model selection, and optimization of jerky type identification.
BibTeX
@inproceedings{han2025non,
  author = {Yuanxi Han and Liang Li and Siyuan Jiang and Pengpeng Sun and Wenliang Wu and Zhendong Liu},
  title = {{Non-destructive identification of commercial jerky types based on multi-band hyperspectral imaging with machine learning}},
  booktitle = {Food},
  year = {2025},
  doi = {10.1016/j.fochx.2025.102293},
}
Figure