基于可见-近红外光谱技术的蜜源快速识别方法

蜂蜜蜜源决定了蜂蜜的药用价值。为了实现快速无损识别蜂蜜蜜源，提出了基于可见-近红外光谱技术结合机器学习的方法来实现蜂蜜蜜源的快速无损识别。该研究采集来自4个蜜源共232份蜂蜜样本光谱数据，随机选取其中212个样本用来构建分类器，剩余20个样本进行分类器泛化学习能力的检验评估。光谱数据预处理采用基线校正，数据标准化和平滑消除干扰和噪声。基于一对多分类规则，采用主成分分析结合贝叶斯线性判别构造线性多分类器，并就分类效果和泛化学习能力与前向神经网络器构成的非线性分类器进行比较。结果表明：基于主成分分析结合贝叶斯线性判别构造的多分类器分类正确率为91.95%，前向神经网络的分类正确率为100%。该研究也表明应用可见-近红外技术对蜂蜜蜜源进行快速分类是可行的。

Rapid recognition method of nectar plant based on visible-near infrared spectroscopy

The honey is worth of nectar plant. A rapid non-destructive method of pattern classification for nectar plant was developed based on visible-near infrared spectroscopy in this study. The nectar plants came from four categories which were Tilia, Astragalus, Leguminosae and Wild hrysanthemum, respectively. A total of 232 samples from four different nectar plants were studied. The calibration set was consisted of 212 samples and the predict set consisted of 20 samples. The classifier was constructed by calibration set which was selected randomly while prediction set was used for evaluating the study ability of classifier. The preprocessing methods were carried on the spectrum data, such as base line correction, normalization and smoothing. The preprocessing methods can enhance signal to noise ratio and remove the random error. The two classifier models were developed using pricipal component analysis combined with Bayesian line discriminant analysis based on one-two-many rule method and backpropagation atificial nerve net method. The accuracy of pricipal component analysis combined with Bayesian line discriminant analysis model was 91.95% and that of the BP- atificial nerve net model was 100%. The results indicated that the nectar plant would be quickly detected by Vis-NIR spectroscopy technique, and it is very feasible.