近红外光谱结合灰狼算法优化支持向量机实现烟叶产地快速鉴别

烟草是一种成分复杂的天然植物，地理位置、生长条件等外界因素直接影响着烟叶的品质；我国烟叶种植范围十分广泛，每个产区种植的烟叶都有其独特的风格特征，不同产区的烟叶配比对卷烟的质量起着决定性的作用。为实现烟叶产地准确、快速判别，基于近红外光谱（NIRS），采用灰狼算法（GWO）优化的支持向量机（SVM）算法实现烟叶产地鉴别分类。以8个产地的824个烟叶样本为研究对象，基于x-y距离样本集划分（SPXY）方法得到校正集617个和验证集207个样品。首先应用最佳波长筛选方法，如竞争自适应加权采样（CARS）和随机青蛙（RF）算法减少光谱冗余信息，最终从1 609个变量中分别获得141和534个与产地相关的重要变量，并以此输入SVM作为建模数据，接下来在相同搜索范围内比较了粒子群优化算法（PSO）、遗传算法（GA）和GWO对SVM分类模型的优化效果。结果表明，经RF筛选后的光谱变量较CARS具有更好的产地建模性能，其中RF-GWO-SVM对8个产地烟叶的整体判别正确率达到了96.62%，相较于RF-PSO-SVM和RF-GA-SVM正确率更高。同时，RF-GWO-SVM的运行时间分别比RF-PSO-SVM和RF-GA-SVM的运行时间缩短156和131 min，RF-GWO-SVM具有精度更高、寻优速度更快等优点。GWO对于SVM模型参数具有更高效的优化能力，可用于烟叶产地快速鉴别模型的建立。

Support Vector Machine Optimized by Near-Infrared Spectroscopic Technique Combined With Grey Wolf Optimizer Algorithm to Realize Rapid Identification of Tobacco Origin

Tobacco is a natural plant with complex compositions, the quality of tobacco leaves is directly affected by several external factors such as geographic location and growth conditions. Tobacco leaves are widely planted in China, and they cultivated in different areas, they have different styles. Different blended ratios play a decisive role in the quality of cigarettes. Thus, there is an emerging need for accurate and rapid identification of the origin of tobacco leaves. Near-infrared spectroscopy technology provides a new rapid, and convenient method to automatically evaluate tobacco areas. On this basis, we proposed the grey wolf optimizer (GWO) algorithm to optimize the performance of the support vector machine model (SVM) for the first time to identify and classify tobacco leaves from different origins. This study was conducted with 824 tobacco leaf samples from eight different origins, and 617 training set samples and 207 test set samples were obtained using Set partitioning based on joint x-y distance (SPXY). The wavelength selection methods such as Competitive adaptive reweighted sampling (CARS) and Random frog (RF) algorithms were applied to reduce spectral redundant information and screen the characteristic wavelengths in the -full spectrum of the samples, and 141 and 534 were selected from all 1 609 variables, respectively. Then they were used as the input parameters of the SVM classifier. The optimization effect of GWO on the SVM model was contrasted to the Particle swarm optimization (PSO) and Genetic algorithm (GA) optimization in the same search range. The analysis showed that the spectral variables screened by RF had a better modeling performance than CARS. Among them, the RF-GWO-SVM model achieved the best predictive performance with an accuracy of 96.62% in identifying tobacco leaves from 8 producing areas. More than that, the running time of RF-GWO-SVM was 156 and 131 min shorter than RF-PSO-SVM and RF-GA-SVM, respectively. To sum up, RF-GWO-SVM has the advantages of higher accuracy and faster convergence speed. It can be seen that GWO has a more efficient optimization capability for model parameters, and the support vector machine model optimized by GWO can be used for rapid identification of tobacco origin.