基于高光谱特征选择的霉变玉米黄曲霉毒素B_1的检测方法

为研究高光谱技术检测霉变玉米中黄曲霉毒素B_1含量的可行性,选择5种不同霉变程度的玉米为试验材料,利用高光谱图像采集系统获得了250个霉变玉米样本的高光谱数据,并进行多元散射校正(MSC)预处理;运用偏最小二乘回归(PLSR)系数来选择特征波长,筛选出7个特征波长,然后利用Fisher判别分析(FDA)分别对全波长和特征波长下霉变玉米进行鉴别分析。结果表明,5组样本在全光谱波段下的FDA鉴别正确率在85%～88%之间,而在特征光谱下的FDA鉴别正确率均在98%以上,说明特征波长能较好地表征不同霉变等级的玉米。神经网络模型优于PLSR模型,其预测集相关系数和均方根误差分别为0.999 9、0.180 9。因此,可认为利用高光谱技术来检测不同霉变程度玉米中的黄曲霉毒素B_1含量是可行的。本研究结果为高光谱鉴别其他农产品提供了重要参考。

Detection Method of Aflatoxin B1 in Moldy Maize Based on Hyperspectral Feature Selection

In order to investigate the feasibility of detecting aflatoxin B₁ in moldy maize using hyperspectral technique, 5 kinds of maize with different moldy degrees were selected as materials. Then the hyperspectral data of the 250 samples were obtained by the hyperspectral image acquisition system, which were preprocessed by multiplicative scatter correction (MSC). The characteristic wavelengths were selected by partial least squares regression (PLSR) coefficients, and 7 characteristic wavelengths were selected. And then Fisher discriminant analysis (FDA) was used to identify the moldy maize samples under full band and characteristic wavelength conditions, respectively. The result showed that the accuracy rate of the 5 kinds of moldy maize samples at the full band was between 85% and 88%, while the accuracy rates of the FDA at these characteristic wavelengths were higher than 98%. This indicated that the different moldy degrees of maize can be characterized by these characteristic wavelengths. The BP model was better than that of PLSR, and correlation coefficient and the root mean square error of the predictive set based on the BP model were 0.999 9 and 0.180 9, respectively. Therefore, it was feasible to detect aflatoxin B₁ content from maize samples with different moldy degree by hyperspectral technology. And an important theoretical reference was also provided for other agricultural products.