【摘要】 不同产地的山楂因生长环境和地理气候的差异，导致其品质的参差不齐，因此确定山楂的地理产区具有重要的意义。为了提高山楂产地溯源模型的稳定性和准确性，提出了一种基于误差倒数赋权的组合鉴别模型。首先，利用高光谱成像技术采集了456个山楂的高光谱信息，通过对比卷积平滑(SG)、多元散射校正(MSC)和标准正态变量(SNV)3种预处理方法，并使用预处理后的数据和原始数据构建BP神经网络(BPNN)和随机森林(RF)模型，根据其准确率确定了SNV为平均光谱值的预处理方法。然后，对山楂的高光谱图像进行主成分分析，选取第1主成分图像并根据全波段下的权重系数筛选出6个特征波长，对应的平均光谱值作为其光谱信息表征值。其次，分别提取第1主成分图像和特征波长对应灰度图像的纹理特征，并将特征波长的光谱表征值与其对应图像的纹理表征值、主成分图像纹理表征值进行组合以构造产地溯源鉴别模型的输入向量。最后，分别选择BP神经网络(BPNN)、随机森林(RF)和赋权组合模型(BPNN-RF)3种方法进行鉴别模型构建，并选取准确率(Acc)和宏F1分数(macroF1)2个评价指标对不同输入向量所构建的山楂产地鉴别模型进行评价分析。结果表明，相同输入向量所建BPNN-RF模型的准确率和宏F1分数大都优于BPNN模型和RF模型的准确率和宏F1分数，其中采用3种表征值组合而成的输入向量，其所建BPNN-RF模型实测集的准确率由89.01%提高到98.90%,宏F1分数也由89.32%提高到98.95%,说明了基于误差倒数赋权BPNN-RF组合模型对山楂产地的鉴别能力最强，效果最好，优于单一的鉴别模型。该研究为不依赖理化分析、仅靠高光谱信息即可实现山楂产地的溯源提供了一种方法支撑。更多还原

【Abstract】 Hawthorns from different origins have uneven quality due to the differences in growth environment and geographic climate, so determining the geographic origin of hawthorns is of great significance. A combined identification model based on error reciprocal weighting was proposed to improve the stability and accuracy of the hawthorn origin traceability model. Firstly, the hyperspectral information of 456 hawthorns was collected using hyperspectral imaging technology; and by comparing Savitzky-Golay Convolutional Smoothing(SG), Multiplicative Scattering Correction(MSC), and Standard Normal Variables(SNV) three preprocessing methods, and used the preprocessed data and the original data to construct BP Neural Network(BPNN) and Random Forest(RF) models, the preprocessing method with SNV as the average spectral value was determined based on their accuracy. Then, the hyperspectral image of the hawthorn was subjected to principal component analysis, and the 1st principal component image was selected; at the same time, six feature wavelengths were screened based on the weight coefficients under the full wavelength band, and then the corresponding average spectral value was used as the representation value of the spectral information. Secondly, the texture features corresponding to the 1st principal component image and the feature wavelengths grayscale images were extracted, respectively, and the spectral representation values of the feature wavelengths were combined with the texture representation values of these feature wavelengths grayscale images as well as the texture representation values of the principal component image to construct the input vectors of the origin traceability identification model. Finally, three methods of BPNN, RF, and weighted combination model(BPNN-RF) were selected for the identification model construction, and two evaluation indexes, namely, accuracy(Acc) and macroF1 score(macroF1) were selected to evaluate and analyze the hawthorn origin identification models constructed by different input vectors. The results showed that the accuracy and macroF1 score of the BPNN-RF model with the same input vector were mostly better than those of the BPNN model and the RF model, in which the accuracy of the actual test data set of the BPNN-RF model with the input vector consisting of three kinds of representation values was increased from 89.01% to 98.90%. The macroF1 score was also increased from 89.32% to 98.95%. This indicates that the combined BPNN-RF model based on the error inverse assignment has the strongest discriminative ability and the best effect on the identification of hawthorn origin, which is better than the single discriminative model such as BPNN or RF. This study provides methodological support for the traceability of hawthorn origin without relying on physicochemical analysis and only relying on hyperspectral information.更多还原