【摘要】 为实现烟田中杂草的精准识别和防控，应用无人机多光谱影像构建光谱、植被、纹理特征等20个特征变量，按照特征重要性评分将特征变量数分为3组，分别添加到随机森林（RF）、支持向量机（SVM）、马氏距离（MD）和C5.0决策树（C5.0）中进行分类，并从分类正确率和时间两方面比较不同分类算法的优适性；通过PCA-ReliefF筛选重要特征信息，在保证较高分类精度的同时降低特征变量维度。结果表明，随着不同特征变量的加入，各分类算法的分类精度不同幅度地提升。其中，使用14个变量的RF模型的分类精度最高、耗时最短，总体精度94.06%,Kappa系数为0.90;MD算法分类精度最低、耗时最少；C5.0决策树耗时最长，总体精度比SVM和MD算法高1.37%、3.62%。综上，RF算法是烟田杂草识别中较合适的分类方法，可以有效实现烟田杂草的精准识别，进而为烟草生产管理提供技术支撑。更多还原

【Abstract】 In order to realize the accurate identification, prevention and control of weeds, UAV multi-spectral images were used to construct 20 characteristic variables, such as spectrum, vegetation and texture features, that were divided into 3 groups according to the importance score of features. The variables were added to random forest(RF), support vector machine(SVM), Markov distance(MD) and C5.0 decision tree(C5.0) for classification. The optimization of different classification algorithms was compared in terms of classification accuracy and time. PCA-relieff was used to screen important feature information and reduce the dimension of feature variables while ensuring high classification accuracy. The results showed that with the addition of different feature variables, the classification accuracy of each classification algorithm increases with different amplitude. Among them, the RF model using 14variables has the highest classification accuracy and the shortest time consuming, with an overall accuracy of 94.06% and a Kappa coefficient of 0.90. MD algorithm has the lowest classification accuracy and the least time consuming. C5.0 decision tree takes the longest time, and its overall accuracy is 1.37% and 3.62% higher than SVM and MD algorithms. In conclusion, RF algorithm is a more suitable classification method for weed identification in tobacco fields, which can effectively realize accurate weed identification, thus provide technical support for tobacco production and management.更多还原