Scholarlink

<p>Soybean quality is crucial for the industry’s competitiveness, but traditional grading methods are slow, subjective, and inaccurate. Non-destructive technologies combined with machine learning algorithms emerge as promising alternatives to overcome these limitations. This study aimed to investigate the use of Near-Infrared (NIR) Spectroscopy and hyperspectral sensors as fast and accurate tools to analyze the physicochemical properties of soybean grains, as well as to identify the machine learning (ML) models with the highest accuracy in classifying the different types. One hundred whole-grain samples were used for each of Types I, II, Basic Standard, and Off-Type. The moisture, crude protein, starch, lipid, crude fiber, and ash contents were determined using a FOSS NIR DS2500 spectrometer. Hyperspectral reflectance data were acquired with an ASD FieldSpec 4 Jr sensor. The instrument covers the 350 to 2500&#xa0;nm spectral range. The acquired spectra were processed and averaged into representative intervals prior to model development. Six ML models were tested in WEKA 3.8.6 softwares: Artificial Neural Networks (ANN), Random Forest (RF), Support Vector Machine (SVM), Zero-R (SL), J48 Decision Tree (J48), and REPTree, evaluated using the metrics correct classification percentage, Kappa, and F-score. Among the models evaluated, the SVM model showed the best performance in classifying the physicochemical quality of soybean grains, achieving 99.5% accuracy, followed by ANN (88.1%) and J48 (76.4%). For the SVM model, the regularization (C) and radial basis kernel (γ) parameters were maintained at the standard values ​​defined by the WEKA 3.8.6 software. The Basic Standard and Off-Type classes obtained the highest class-specific accuracies among all grain types. The integration of NIR, hyperspectral sensors, and ML enabled the characterization of the physicochemical quality of the soybeans, demonstrating classification efficiency and outperforming conventional methods. The average analysis time per sample was approximately one minute, highlighting the potential of this approach to automate soybean grading, with greater reliability, agility, and standardization of the process.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A method for classifying the physicochemical quality of soybean grains using near-infrared spectroscopy, hyperspectral sensor technology, and artificial intelligence models

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月6日

查看全文
  • Marisa Menezes Leal,
  • Nairiane dos Santos Bilhalva,
  • Rosana Santos de Moraes,
  • Dthenifer Cordeiro Santana,
  • Larissa Pereira Ribeiro Teodoro,
  • Paulo Eduardo Teodoro,
  • Paulo Carteri Coradi

摘要

Soybean quality is crucial for the industry’s competitiveness, but traditional grading methods are slow, subjective, and inaccurate. Non-destructive technologies combined with machine learning algorithms emerge as promising alternatives to overcome these limitations. This study aimed to investigate the use of Near-Infrared (NIR) Spectroscopy and hyperspectral sensors as fast and accurate tools to analyze the physicochemical properties of soybean grains, as well as to identify the machine learning (ML) models with the highest accuracy in classifying the different types. One hundred whole-grain samples were used for each of Types I, II, Basic Standard, and Off-Type. The moisture, crude protein, starch, lipid, crude fiber, and ash contents were determined using a FOSS NIR DS2500 spectrometer. Hyperspectral reflectance data were acquired with an ASD FieldSpec 4 Jr sensor. The instrument covers the 350 to 2500 nm spectral range. The acquired spectra were processed and averaged into representative intervals prior to model development. Six ML models were tested in WEKA 3.8.6 softwares: Artificial Neural Networks (ANN), Random Forest (RF), Support Vector Machine (SVM), Zero-R (SL), J48 Decision Tree (J48), and REPTree, evaluated using the metrics correct classification percentage, Kappa, and F-score. Among the models evaluated, the SVM model showed the best performance in classifying the physicochemical quality of soybean grains, achieving 99.5% accuracy, followed by ANN (88.1%) and J48 (76.4%). For the SVM model, the regularization (C) and radial basis kernel (γ) parameters were maintained at the standard values ​​defined by the WEKA 3.8.6 software. The Basic Standard and Off-Type classes obtained the highest class-specific accuracies among all grain types. The integration of NIR, hyperspectral sensors, and ML enabled the characterization of the physicochemical quality of the soybeans, demonstrating classification efficiency and outperforming conventional methods. The average analysis time per sample was approximately one minute, highlighting the potential of this approach to automate soybean grading, with greater reliability, agility, and standardization of the process.