产地是影响农作物生产的重要环境因素, 产地溯源对于食品安全具有重要意义。 针对传统农产品产地检测一般采用化学分析法, 其操作繁琐且存在破坏性和耗时较长等不足的问题, 以北方寒地绿豆为研究对象, 分别在白城、 杜蒙、 泰来等优质绿豆主产区, 获取绿豆的籽粒和粉末两种状态的近红外光谱数据（NIR）, 利用优选NIR光谱特征波数, 建立了绿豆产地无损检测的新方法。 首先在吸光度值较强的10 105.37~4 078.655 cm-1波数范围内, 采用多元散射校正法（MSC）, 对不同产地的绿豆原始光谱数据进行预处理, 以消除光谱干扰信息。 应用竞争性自适应重加权采样算法（CARS）, 优选不同产地绿豆籽粒和粉末状态的光谱特征波数, 以减少光谱曲线的特征向量维度。 最后利用前馈神经网络（BP）自适应推理机制, 建立了绿豆产地与其光谱特征波数之间非线性映射模型, 并将网络输出的编码向量解析至产地名称, 作为绿豆产地检测的输出结果。 研究结果表明: （1）原始光谱经过多元散射校正预处理后, 绿豆粉末光谱曲线的误差从12.87降到3.20, 绿豆籽粒光谱曲线的误差从153.04降到27.73, 提供有效可靠的光谱数据。 （2）通过竞争性自适应重加权采样算法, 提取绿豆光谱曲线的重要特征波数, 从籽粒和粉末状态原始2 114个波数中, 分别优化为61个和107个特征波数, 波段总数目减少了94.94%以上, 并将其作为绿豆产地识别的特征指标。 （3）创新性提出了MSC-CARS-BP绿豆产地检测模型, 以优选出的光谱特征波数为定量依据, 分别对绿豆籽粒和粉末进行产地检测, 预测集准确率为92.59%和98.63%, 相关系数均达到0.99以上。 该方法能够利用近红外光谱处理技术, 实现绿豆产地无损检测的目标, 为农产品产地自动快速溯源提供了技术支持和参考。

Origin is an important environmental factor affecting crop production, and tracing the origin is of great significance for food safety. The chemical analysis method is generally used in traditional agricultural product origin detection, and its operation is cumbersome, destructive and time-consuming. In this study, northern cold mung beans were used as the research object. Near-infrared spectral data of mung bean in two states of seed and powder were obtained in the main origins of high-quality for Baicheng, Dumeng and Tailai. A new nondestructive detecting method for mung bean origins were established by optimizing the NIR characteristic spectrum wavenumbers. Firstly, in the range of 10 105.37~4 078.655 cm-1 wavenumber with strong absorbance value, the raw spectral data of mung beans from different regions was preprocessed by using multivariate scattering correction (MSC) method to eliminate spectral interference information. Then competitive adaptive reweighted sampling(CARS) algorithm is applied to optimize the characteristic spectral wavenumbers of mung bean seed and powder states from different origins to reduce the feature vector dimension of the spectral curve. Finally, a feed-forward neural network (BP) adaptive inference mechanism was used to establish a non-linear mapping model between the origin of mung bean and its spectral characteristic wavenumber, and the encoding vector output by the network was parsed to the original name as the output result of the detection of the origin of the mung bean. The results show that: (1)Preprocessed with multiple scattering corrections in the raw spectral, the error of the spectral curve of mung bean powder is reduced from 12.87 to 3.20, and the error of the spectral curve of mung bean seed is reduced from 153.04 to 27.73, which provides effective and reliable spectral data. (2) Through the competitive adaptive reweighting sampling algorithm, the important characteristic wavenumbers of mung bean spectral curve are extracted. From the 2 114 original wavenumbers of seed and powder state, 61 and 107 characteristic wavenumbers are optimized respectively, and the total number of wavebands is reduced by 94.94%, which is taken as the characteristic index of mung bean origin recognition. (3) The MSC-CARS-BP mung bean origin detection model was put forward innovatively. Based on the optimized spectral characteristic wavenumber as the quantitative basis, the origin detection of mung bean seed and powder was carried out respectively. The accuracy of the prediction set was 92.59% and 98.63%, and the correlation coefficient was above 0.99. This method can use near-infrared spectrum processing technology to achieve the goal of non-destructive detecting of mung bean origin, and provide technical support and reference for automatic and rapid traceability of agricultural products origin.