Abstract: To achieve rapid and non⁃destructive monitoring of chemical quality indicators in tobacco，this study was conducted at the Tangyin Tobacco Station in Fuzhou City，Jiangxi Province，using two tobacco varieties，Yunyan 87 and ZY157，under five nitrogen application levels（0，105.0，127.5，157.5，and 180. 0 kg/ha）. During key growth stages，multispectral and RGB images along with three⁃dimensional point cloud data were acquired using DJI Phantom 4 and Mavic 3T drones. Spectral vegetation indices，color indices，texture features，and three⁃dimensional structural features were extracted from the data.Estimation models for key chemical components such as nicotine，nitrogen，and potassium were constructed using random forest（RF），least absolute shrinkage and selection operator（LASSO），and partial least square regression（PLSR）algorithms. The results indicated that all feature types exhibited strong responsiveness to tobacco quality，with the red⁃edge optimized soil⁃adjusted vegetation index（REOSAVI） showing the higher correlation with nicotine，nitrogen，and potassium contents（|r| ranged from 0.68 to 0.86）. Compared to single⁃feature modeling，multi⁃source feature fusion modeling significantly improved prediction accuracy. For nitrogen content monitoring，the texture feature combination performed best（R²=0. 89），while for nicotine and potassium content prediction，the fusion of all four feature types achieved optimal results，with R² values of 0.88—0.89 and RMSE values of 0.22%—0.31%. Algorithm comparison revealed that RF and LASSO outperformed PLSR in both fitting and predictive capabilities，with LASSO demonstrating the best performance in multi⁃feature fusion modeling. Spatial distribution maps of quality parameters generated based on the optimal model clearly reflected the spatial variability of tobacco quality under different nitrogen application levels. In conclusion，integrating multi⁃source UAV remote sensing features with machine learning algorithms effectively enhances the estimation accuracy of tobacco chemical quality indicators，providing technical support for tobacco quality monitoring，precision fertilization，and intelligent field management.

Key words: Tobacco quality, Unmanned aerial vehicle（UAV）, Multi?source remote sensing, Feature fusion, Machine learning, Non?destructive monitoring

摘要： 为实现烟草关键化学品质指标的高效无损监测，以江西省抚州市棠阴烟站为试验地点，以云烟87与ZY157品种为试验材料，设置5个施氮水平（0、105.0、127.5、157.5、180.0 kg/hm²）。于烟草关键生育时期，利用大疆精灵4与Mavic 3T无人机分别获取多光谱、RGB影像及三维点云数据，并从中提取光谱植被指数、颜色指数、纹理特征与三维结构特征。基于随机森林（RF）、最小绝对收缩和选择算子（LASSO）回归及偏最小二乘回归（PLSR）算法，构建烟草烟碱、氮、钾等化学成分的无损监测模型。结果表明，红边优化土壤调节植被指数（REOSAVI）与烟碱、氮、钾含量具有更高的相关性，相关系数（r）绝对值为0.68~0.86。相较于单一特征模型，多源特征融合模型显著提升了模型预测精度。在氮含量监测中，纹理特征组合表现最优（R^²=0.89）；而对于烟碱与钾含量的监测，4类特征融合模型效果最佳，R^²达0.88~0.89，均方根误差（RMSE）为0.22%~0.31%。模型算法对比显示，RF与LASSO在拟合与监测能力上均优于PLSR，其中，LASSO在多特征融合建模中表现最为稳健。基于最优模型绘制的品质空间分布图，清晰地揭示了不同施氮处理下烟草化学品质的空间分异规律。综上，融合多源无人机遥感特征并结合机器学习算法构建模型，可有效提升烟草化学品质指标的监测精度，为烟草品质动态监测、精准施肥与田间智慧管理提供可靠的技术支持。

关键词: 烟草品质, 无人机, 多源遥感, 特征融合, 机器学习, 无损监测