Scholarlink

<p>Accurate and non-destructive estimation of wheat protein content is critical for quality-based grain management, precision agriculture, and food industry applications. This study evaluates the performance of traditional and advanced machine learning approaches for predicting wheat flour protein content using spectroradiometric reflectance data across the 350–2500&#xa0;nm range. Partial Least Squares Regression (PLSR), Random Forest (RF), Deep Learning (DL), and a hybrid Artificial Neural Network–Support Vector Regression (ANN–SVR) model were systematically compared using region-specific spectral subsets, including the visible (VIS: 350–700&#xa0;nm), near-infrared (NIR: 700–1250&#xa0;nm), short-wave infrared (SWIR: 1250–2500&#xa0;nm), and the full spectrum. The results revealed clear performance differences among modeling approaches. PLSR showed limited predictive capability across all spectral regions, indicating its restricted ability to capture non-linear relationships in high-dimensional spectral data. RF models provided improved accuracy by exploiting ensemble learning and feature selection mechanisms, although their performance varied depending on spectral domain. The highest and most stable prediction accuracy was achieved by the ANN–SVR hybrid and deep learning models, which effectively modeled complex non-linear interactions between protein content and spectral reflectance. Among spectral domains, NIR and SWIR regions exhibited strong sensitivity to protein-related absorption features, while the full-spectrum models consistently delivered the best overall performance by integrating complementary spectral information. Spectral difference analysis and principal component analysis further confirmed that protein-related variability is primarily associated with N–H and C–H overtone absorption bands in the NIR–SWIR transition zones. The findings demonstrate that combining hyperspectral spectroscopy with hybrid and deep learning architectures substantially enhances prediction accuracy compared to conventional linear models. Overall, this study highlights the potential of hyperspectral data-driven artificial intelligence models as a rapid, non-destructive, and scalable alternative to conventional laboratory protein analysis. The proposed framework supports real-time quality monitoring, site-specific nitrogen management, and post-harvest grain classification, contributing to sustainable agricultural practices and digital transformation in wheat quality assessment.</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

Estimation of protein content in wheat using traditional regression, hybrid models, and deep learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月7日

查看全文
  • Süreyya Betül Rufaioğlu,
  • Ali Volkan Bilgili,
  • Amjad Mohamed Ismael,
  • Yüsra İnci

摘要

Accurate and non-destructive estimation of wheat protein content is critical for quality-based grain management, precision agriculture, and food industry applications. This study evaluates the performance of traditional and advanced machine learning approaches for predicting wheat flour protein content using spectroradiometric reflectance data across the 350–2500 nm range. Partial Least Squares Regression (PLSR), Random Forest (RF), Deep Learning (DL), and a hybrid Artificial Neural Network–Support Vector Regression (ANN–SVR) model were systematically compared using region-specific spectral subsets, including the visible (VIS: 350–700 nm), near-infrared (NIR: 700–1250 nm), short-wave infrared (SWIR: 1250–2500 nm), and the full spectrum. The results revealed clear performance differences among modeling approaches. PLSR showed limited predictive capability across all spectral regions, indicating its restricted ability to capture non-linear relationships in high-dimensional spectral data. RF models provided improved accuracy by exploiting ensemble learning and feature selection mechanisms, although their performance varied depending on spectral domain. The highest and most stable prediction accuracy was achieved by the ANN–SVR hybrid and deep learning models, which effectively modeled complex non-linear interactions between protein content and spectral reflectance. Among spectral domains, NIR and SWIR regions exhibited strong sensitivity to protein-related absorption features, while the full-spectrum models consistently delivered the best overall performance by integrating complementary spectral information. Spectral difference analysis and principal component analysis further confirmed that protein-related variability is primarily associated with N–H and C–H overtone absorption bands in the NIR–SWIR transition zones. The findings demonstrate that combining hyperspectral spectroscopy with hybrid and deep learning architectures substantially enhances prediction accuracy compared to conventional linear models. Overall, this study highlights the potential of hyperspectral data-driven artificial intelligence models as a rapid, non-destructive, and scalable alternative to conventional laboratory protein analysis. The proposed framework supports real-time quality monitoring, site-specific nitrogen management, and post-harvest grain classification, contributing to sustainable agricultural practices and digital transformation in wheat quality assessment.