摘要
粮食霉变引发的真菌毒素生物污染是威胁粮食供应链安全的重大风险源,建立早期霉变预警机制对保障粮食品质具有关键防控价值。伴随着光学技术和检测技术的进一步发展,无损的光谱成像技术成为粮食霉变检测的热点。传统光谱成像技术存在图像采集速度慢、时间成本高的问题,相比之下,机器学习辅助光谱成像展现出了更快的检测速度和更高的检测准确率,因而受到广泛关注。而不同的机器学习算法与不同的光谱成像技术有着差异化的契合度。本文聚焦机器学习辅助光谱成像的粮食霉变检测,系统综述了该交叉领域的技术突破与理论进展,探究了高光谱成像、太赫兹时域光谱等多类型光谱技术的原理与特性,分析了多种机器学习算法在光谱成像中的关键作用,包括研究多种光谱预处理方法(如一阶导数、基线校正、SG平滑)对数据信噪比的提升作用,同时也比较了机器学习中监督和无监督机器学习算法对不同的高光谱数据产生的影响,对机器学习算法的优化和光谱技术在粮食霉变无损检测体系的构建提出了建议。
Mycotoxin contamination caused by grain mildew poses a significant risk to global grain supply chain safety,and establishing early mildew warning mechanisms is critical for safeguarding grain quality.With advancements in optical and detection technologies,nondestructive spectral imaging has emerged as a prominent approach for grain mildew detection.Traditional spectral imaging techniques suffer from slow image acquisition speeds and high time costs,whereas machine learning(ML)-assisted spectral imaging has garnered widespread attention due to its superior detection speed and accuracy.Notably,distinct ML algorithms exhibit differentiated compatibility with various spectral imaging modalities.This review focuses on ML-augmented spectral imaging for grain mildew detection,systematically summarizing technological breakthroughs and theoretical progress in this interdisciplinary field.The principles and characteristics of hyperspectral imaging(HSI)and terahertz time-domain spectroscopy(THz-TDS),and the pivotal roles of ML algorithms in spectral data processing are analyzed.Key aspects include the efficacy of spectral preprocessing methods-such as first-order derivative transformation,baseline correction,and Savitzky-Golay(SG)smoothing-in enhancing data signal-to-noise ratios(SNR).Furthermore,the impacts of supervised and unsupervised ML algorithms on hyperspectral data interpretation are compared.Optimization strategies for ML algorithms and spectral technologies are proposed to advance nondestructive mildew detection systems.
作者
何俊杰
何阳
禹唯
鲁玉杰
关桦楠
He Junjie;He Yang;Yu Wei;Lu Yujie;Guan Huanan(School of Grain Science and Technology,Jiangsu University of Science and Technology,Zhenjiang 212003)
出处
《中国粮油学报》
北大核心
2025年第11期29-40,共12页
Journal of the Chinese Cereals and Oils Association
基金
国家重点研发计划项目(2023YFC2604903)
国家自然科学基金项目(32372417)
黑龙江省自然科学基金资助项目(LH2022C046)
黑龙江省博士后科研启动项目(LBH-Q19027)
黑龙江省领军人才支持计划项目(2020376)
中央财政支持地方高校发展专项(YSL036)。
关键词
粮食霉变
高光谱成像
太赫兹光谱
机器学习
识别
检测
grain mildew
hyperspectral imaging
terahertz spectroscopy
machine learning
recognition
detection
