Scholarlink

<p>Accurate determination of crude fat content in maize kernels is essential for quality grading and utilization, yet conventional physicochemical methods are destructive, time-consuming, and unsuitable for large-scale or real-time applications. Existing NIR-based approaches are often limited by inadequate feature representation and insufficient nonlinear regression. To address these challenges, a rapid and non-destructive analytical framework was developed by integrating NIR spectroscopy with a dual-attention residual deep learning architecture and a Kolmogorov–Arnold Network–based prediction head. This design enables efficient extraction of chemically relevant spectral features and robust nonlinear regression. Experimental results demonstrated that the proposed model achieved the best performance among the evaluated models on an independent test set for maize crude fat prediction, with an Rp of 0.88, RMSE of 0.33, and RPD of 2.09. The SHAP analysis revealed that the model focused on lipid-related C–H overtone and combination band regions, supporting the chemical interpretability of the predictions. Overall, this study indicates that NIR spectroscopy combined with interpretable deep learning can provide a practical decision-support tool for rapid, non-destructive screening of crude fat content in maize kernels.</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

Non-destructive Quantification of Crude Fat in Maize Kernels Using Near-infrared Spectroscopy and a Dual-attention Residual Kolmogorov–Arnold Network

  • 期刊论文

  • 发表日期：

    2026年7月28日

查看全文
  • Lanxiang Yu,
  • Shiyang Wei,
  • Hualong Xu,
  • Yaqi Hu,
  • Caihong Wang,
  • Mingxing Li,
  • Yao Qin,
  • Xueqin Wei,
  • Yu Yang

摘要

Accurate determination of crude fat content in maize kernels is essential for quality grading and utilization, yet conventional physicochemical methods are destructive, time-consuming, and unsuitable for large-scale or real-time applications. Existing NIR-based approaches are often limited by inadequate feature representation and insufficient nonlinear regression. To address these challenges, a rapid and non-destructive analytical framework was developed by integrating NIR spectroscopy with a dual-attention residual deep learning architecture and a Kolmogorov–Arnold Network–based prediction head. This design enables efficient extraction of chemically relevant spectral features and robust nonlinear regression. Experimental results demonstrated that the proposed model achieved the best performance among the evaluated models on an independent test set for maize crude fat prediction, with an Rp of 0.88, RMSE of 0.33, and RPD of 2.09. The SHAP analysis revealed that the model focused on lipid-related C–H overtone and combination band regions, supporting the chemical interpretability of the predictions. Overall, this study indicates that NIR spectroscopy combined with interpretable deep learning can provide a practical decision-support tool for rapid, non-destructive screening of crude fat content in maize kernels.