基于高光谱成像的玉米渣中伏马菌素快速无损检测模型研究

Study on rapid and nondestructive detection model of fumonisin in corn residue based on hyperspectral imaging

中文关键词:  玉米渣  伏马毒素（FUMs）  高光谱成像  深度学习

英文关键词:corn residue  fumonisins  hyperspectral imaging  interpretable deep learning

中文摘要:

      目的 探索采用高光谱成像技术(hyperspectral imaging, HSI)(光谱范围900~1700 nm)结合深度学习方法构建玉米渣中伏马菌素(fumonisin, FUMs)快速无损检测模型的可行性。方法 通过人工制备含不同浓度梯度FUMs (0~6 mg/kg)的玉米渣样品(共175个), 经多元散射校正(multiplicative scatter correction, MSC)预处理消除干扰后, 采用无信息变量消除法(uniformative variable elimination, UVE)与变量组合群体分析法(variable combination population analysis, VCPA)筛选出26个关键特征波长。结果 对比偏最小二乘回归(partial least squares regression, PLSR)、支持向量机(support vector machine, SVM)、卷积神经网络(convolutional neural network, CNN)、长短期记忆网络(long short-term memory, LSTM)、门控循环单元(gated recurrent unit, GRU)和注意力机制(Attention)等模型后, 发现CNN-GRU-Attention融合模型表现最优: FUMs检测的预测决定系数达0.9328、预测均方根误差为0.3931、预测偏差比为3.8247。结论 本研究建立的检测方法实现了FUMs的快速、精准、无损同步检测, 为谷物真菌毒素检测提供了新的技术路径, 也为深度学习在光谱分析领域的应用提供了参考。

英文摘要:

      Objective To explore the feasibility of constructing a rapid and nondestructive detection model for fumonisin (FUMs) in corn residue using hyperspectral imaging (HSI, spectral range 900–1700 nm) combined with deep learning methods. Methods Corn bran samples containing a concentration gradient of FUMs (0–6 mg/kg) were artificially prepared (total of 175 samples). After removing interference through multiplicative scatter correction (MSC) preprocessing, the univariate variable elimination (UVE) and variable combination population analysis (VCPA) methods were used to identify key feature wavelengths for FUMs (26 wavelengths), respectively. Results After comparing models including partial least squares regression (PLSR), support vector machine (SVM), convolutional neural network (CNN), long short-term memory (LSTM), gated recurrent unit (GRU) and Attention, the CNN-GRU-Attention fusion model demonstrated optimal performance: For FUMs detection, the prediction coefficient of determination reached 0.9328, the root mean square prediction error was 0.3931, and the prediction deviation ratio was 3.8247. Conclusion The established detection method enables rapid, precise and nondestructive simultaneous detection of FUMs, providing a novel technical pathway for grain mycotoxin detection and offering reference for the application of interpretable deep learning in spectral analysis.