定量分析(化学)
融合
光谱学
生物系统
化学
计算机科学
材料科学
分析化学(期刊)
定性分析
定量评估
传感器融合
融合蛋白
定量蛋白质组学
分析技术
人工智能
数据挖掘
模式识别(心理学)
工艺工程
统计分析
信号处理
色谱法
作者
Xiao Dong Xie,Tian Jiao Yang,Kevin Fynn,Yang Li
标识
DOI:10.1080/10739149.2026.2681071
摘要
Wheat protein content is an important indicator of wheat quality. Conventional analytical methods, such as the Kjeldahl method, are accurate but labor-intensive and time-consuming. Near-infrared spectroscopy (NIRS) provides a rapid and nondestructive alternative with simple sample preparation. However, calibration models developed on one instrument often perform poorly on another instrument because of systematic spectral differences, which limits the broader application of NIRS in grain quality assessment. In this study, a multi-scale attention fusion network (MSAF-Net) was developed for the quantitative prediction and cross-instrument transfer of wheat protein content. The model integrates channel–wavelength attention, dynamic multi-scale feature fusion, and domain-adversarial transfer to improve feature extraction and reduce the influence of instrument-related spectral variation. Experiments on the IDRC 2016 wheat protein spectral dataset showed that MSAF-Net achieved a coefficient of determination (R2) of 0.9473 and a root mean square error (RMSE) of 0.0340 on the primary instrument, corresponding to a 16.9% reduction in RMSE compared with partial least squares (PLS). In cross-instrument transfer, fine-tuning with 50 target-instrument samples achieved an R2 of 0.9774 and reduced RMSE by 63.2% compared with the traditional PLS transfer strategy. With 0.026 million trainable parameters and 14.258 M MACs for single-sample inference, the model also showed potential for deployment on portable NIR instruments. These results indicate that MSAF-Net can improve the accuracy and transferability of NIRS-based wheat protein prediction.
科研通智能强力驱动
Strongly Powered by AbleSci AI