高光谱最优波长选择及Fisher判别分析法判别玉米颗粒表面黄曲霉毒素

摘要

Aflatoxin is a toxic metabolite widely existing in corn .In the present paper ,the feasibility of detecting aflatoxin on corn kernel surface by hyperspectral imaging technology was verified .The corn called pioneer with the same shape is provided by Toxicology and Mycotoxin Research Unit .With methanol configuration ,four different concentrations of aflatoxin solutions were prepared and dripped on every 30 corn kernels .Also other clean 30 kernels without aflatoxin dripped were prepared to be the control samples .Among the 150 kernel samples ,103 training samples and 47 validation samples were prepared randomly .First-ly ,hyperspectral image in the range of 400 to 1 000 nm was collected .For eliminating the deviations in original spectrum ,stand-ard normal variate transformation (SNV) was adopted as pretreatment method .And then several optimum wavelengths were se-lected by the principle of minimum misdiagnosis rate .After that the selected optimum wavelengths were taken as the input of the Fisher discrimination analysis to discriminate the different concentrations of aflatoxin on the corn .Finally ,the discrimination model based on four optimum wavelengths (812.42 ,873.00 ,900.36 and 965.00 nm) was built and the accuracy of the model was tested .Results indicate that the classification accuracy of calibration and validation set was 87.4% and 80.9% respectively . This method provides basis for designing the corresponding portable instrument and distinguishing aflatoxin produced by natural-ly metabolism in corn .%黄曲霉毒素是广泛存在于玉米中且具有剧毒的一种代谢产物，以美国农业部农业研究署（USDA-ARS） Toxicology and Mycotoxin Research Unit提供的2010年先锋玉米为研究对象，验证了高光谱成像技术对玉米中黄曲霉毒素检测的可行性。以甲醇为溶剂制备四种不同浓度的黄曲霉毒素溶液，并将其逐一滴在等量的4组共120粒玉米颗粒表面，以未处理的30粒洁净玉米作为一组对照样本，将大小、形状相似的150个样品随机分为训练集103个，验证集47个；对获取的400～1000 nm波段范围内的高光谱图像，先进行标准正态变量变换（standard normal variate transformation ，SNV）预处理，然后引入基于 Fisher判别最小误判率的方法选择最优波长，并以所选波长作为Fisher判别分析法的输入建立判别模型，对玉米颗粒表面不同浓度的黄曲霉毒素进行识别，最后对模型判别正确率进行了验证。结果表明，选取四个最优波长（812.42，873.00，900.36和965.00 nm ）时Fisher判别分析模型对训练集与验证集的准确率分别为87.4％和80.9％。该方法为含黄曲霉毒素玉米颗粒便携式检测仪器的开发，以及对田间霉变玉米自然代谢产生毒素的检测奠定了技术基础。