Abstract:

To rapidly and accurately obtain the pesticide types used in tobacco field production， and to improve the scientific and targeted management of tobacco field pesticide applications， hyperspectral imaging technology was applied. By comparing the spectral curves of fresh tobacco leaves 48 h after spraying with 5 common pesticides， the spectral differences were analyzed. Various combinations of spectral preprocessing methods， feature wavelength extraction methods， and pattern recognition techniques were tested to evaluate model accuracy. The results showed that， in the 750~875 nm spectral range， the spectral curves of fresh tobacco leaves treated with the 5 pesticides exhibited distinct reflectance differences. Across the full spectral range， the combination of standard normal variate （SNV） transformation and least squares support vector machine （LSSVM） models， as well as the combination of second derivative preprocessing and random forest models， both achieved high recognition accuracy， with the test set accuracy reaching 98.58%. The continuous projection algorithm outperformed the competitive adaptive reweighting sampling algorithm in dimensionality reduction. In the feature wavelength range， the combination of second derivative preprocessing， continuous projection algorithm， and random forest models achieved the best performance， with the training set accuracy reaching 100.00% and the test set accuracy at 98.22%. The number of feature wavelengths was 17， and the single-sample detection time was 10.28 ms. This method could rapidly and accurately identify the types of 5 common pesticides used in tobacco production. Above results provided technical support for the management of pesticide application in tobacco field production.

Key words: tobacco field production, hyperspectral imaging, pesticide identification, random forest

摘要：

为快速、准确获取烟叶生产田间用药类型，提升烟叶田间用药管控科学性和针对性，应用高光谱成像技术，通过对比分别喷施5种烟草常用农药48 h后的鲜烟叶光谱曲线差异，测试不同光谱预处理方法、特征波长提取方法和模式识别方法组合下的模型准确率。结果表明，在750~875 nm波段，5种农药处理后的鲜烟叶光谱曲线反射率存在差异。全波段下，标准正态变换和最小二乘支持向量机模型组合以及二阶导数和随机森林模型组合均具有很高的识别准确率，测试集准确率均达到98.58%；连续投影算法降维效果优于竞争性自适应重加权采样算法。特征波段下，二阶导数、连续投影算法和随机森林的模型组合表现最优，训练集识别准确率为100.00%，测试集识别准确率为98.22%，特征波段数量为17，单样本检测时间为10.28 ms，该方法可快速、准确地识别5种烟草生产常用农药类型。研究结果为烟叶生产田间用药管控提供技术支持。

关键词: 烟叶生产, 高光谱成像, 农药识别, 随机森林