近红外光谱技术快速无损检测轻度霉变玉米

Rapid non-destructive detection of mildly moldy Zea mays by near-infrared spectroscopy technology

中文关键词:  玉米  近红外光谱  黄曲霉毒素  无损检测

英文关键词:Zea mays  near-infrared spectroscopy  aflatoxin  backpropagation neural network

中文摘要:

      目的 应用近红外光谱(near-infrared spectroscopy, NIRS)技术对轻度霉变玉米中黄曲霉毒素进行快速无损检测。方法 本研究选取轻度霉变的玉米样本作为实验材料, 以黄曲霉毒素B1 (aflatoxin B1, AFB1)的含量作为检测指标, 利用NIRS图像采集系统收集了153个样本图像, 采用多元散射校正、标准正态变换及移动平均平滑(moving average smoothing, MAS) 3种预处理方法对样本的原始NIRS数据(raw near-infrared spectral data, RNSD)进行预处理。并采用反向传播神经网络(backpropagation neural network, BPNN)、极限学习机和支持向量机对处理后的光谱数据与AFB1含量数据进行建模分析, 评估预处理方法对模型性能的影响; 并通过连续投影算法(stepwise projection algorithm, SPA)对预处理数据进行特征光谱选取后代入模型进行综合比较。结果 最佳光谱预处理方法为MAS, 通过SPA选取出10个特征光谱, 使用BPNN模型时预测结果最佳, 模型预测集的决定系数可达到0.932, 相对预测偏差为3.922, 该模型具有良好的性能和可靠性。结论 利用NIRS技术测定轻度霉变玉米中的AFB1含量具备可行性, 本研究成果为NIRS在鉴别其他农产品方面的应用提供了重要的参考依据。

英文摘要:

      Objective To rapidly and non-destructively detect aflatoxin in mildly moldy Zea mays using near-infrared spectroscopy (NIRS) technology. Methods Mildly moldy Zea mays samples were selected as experimental materials, with the content of aflatoxin B1 (AFB1) as the detection indicator. A total of 153 sample images were collected using the NIRS imaging acquisition system. Three kinds of preprocessing methods, including multiplicative scatter correction, standard normal variate transformation, and moving average smoothing (MAS), were applied to preprocess the raw near-infrared spectral data (RNSD). Backpropagation neural network (BPNN), extreme learning machine, and support vector machine were employed to model and analyze the preprocessed spectral data along with AFB1 content data, evaluating the impact of preprocessing methods on model performance. Furthermore, the stepwise projection algorithm (SPA) was performed to select characteristic spectra from the preprocessed data for comprehensive comparison after incorporating them into the models. Results The optimal spectral preprocessing method was MAS. Ten characteristic spectra were selected through SPA, and the BPNN model exhibited the best prediction results, achieving a coefficient of determination of 0.932 and a relative prediction deviation of 3.922. This model demonstrated good performance and reliability. Conclusion It is feasible to determine AFB1 content in mildly moldy Zea mays using NIRS technology. The findings of this study provide an important reference for the application of NIRS in identifying other agricultural products.