基于无人机多光谱图像和集成学习的橡胶树白粉病检测  

Detection of rubber tree powdery mildew based on UAV multispectral image and stacked machine learning model

作　　者：王勇 曾体伟 徐秋 付威 付梦 张慧明 WANG Yong;ZENG Tiwei;XU Qiu;FU Wei;FU Meng;ZHANG Huiming(Mechanical and Electrical Engineering College,HainanUniversity,Haikou,Hainan 570228,China;School of Information and Communication Engineering,HainanUniversity,Haikou,Hainan 570228,China)

机构地区：[1]海南大学机电工程学院,海南海口570228 [2]海南大学信息与通信工程学院,海南海口570228

出　　处：《石河子大学学报（自然科学版）》2024年第3期265-274,共10页Journal of Shihezi University(Natural Science)

基　　金：海南省自然科学基金项目(521RC1036),国家自然科学基金项目(32160424)。

摘　　要：橡胶树是我国重要的热带经济作物,其生长过程中易受白粉病的侵染,准确、及时地监测橡胶树白粉病是防止其大规模蔓延的关键。近年来,无人机遥感技术在农林领域得到了广泛应用,本研究评估了采用低空遥感技术大规模检测橡胶树白粉病的可行性,并致力于提高检测的准确性。基于大疆精灵4多光谱无人机获取橡胶树冠层多光谱图像,计算植被指数(VI)和纹理特征(TF),然后结合皮尔逊相关系数(PCCs)和Boruta-SHAP算法进行相关性分析和特征重要性分析,去除冗余特征,Blue-MEA、WI、DVIRE、PPR和GI被选为最佳特征组合,最后基于K近邻(KNN)、朴素贝叶斯(Bayes)、支持向量机(SVM)、随机森林(RF)、极端梯度提升(XGB)和Stacking集成算法构建橡胶树白粉病监测模型。结果表明:经特征筛选后,Stacking集成模型的准确率(OA)和Kappa(KC)值分别达到96.39%和92.78%,相比于5个单一基础模型KNN、Bayes、SVM、RF、XGB分类的效果,集成学习模型的准确率分别提高了3.15%、5.52%、1.80%、3.04%、1.14%,Kappa值提高了6.32%、11.05%、3.61%、6.09%、2.27%;在绘制橡胶树白粉病空间分布图时,使用17×17窗口大小的像素聚合策略分类准确率最高(OA=96.22%)。Rubber trees are a significant tropical cash crop in China that are vulnerable to powdery mildew.Monitoring this disease accurately and promptly is crucial in preventing widespread infestations.Recently,UAV remote sensing technology has been increasingly utilized in agriculture and forestry.This study aimed to evaluate the potential of low-altitude remote sensing technology in detecting powdery mildew in rubber trees on a large scale and enhancing detection accuracy.The process involved acquiring multispectral images of rubber tree canopies using a DJI Phantom 4 multispectral drone,calculating vegetation index(VI)and texture features(TF),conducting correlation and feature importance analyses with Pearson correlation coefficient(PCCs)and Boruta-SHAP algorithms to eliminate redundant features,and selecting Blue-MEA,WI,DVIRE,PPR,and GI as the optimal feature combinations.Subsequently,a monitoring model for rubber tree powdery mildew was developed using K-nearest neighbor(KNN),naive Bayes,support vector machine(SVM),random forest(RF),extreme gradient boosting(XGB),and Stacking integration algorithms.The results indicated that after feature selection,the Stacking integrated model achieved an accuracy(OA)and Kappa(KC)value of 96.39%and 92.78%,respectively.The integrated learning model showed a 3.15%,5.52%and 1.80%improvement in accuracy compared to the individual base models(KNN,Bayes,SVM,RF,and XGB),with Kappa values increasing by 6.32%,11.05%,3.61%,6.09%and 2.27%.Notably,the highest classification accuracy(OA=96.22%)was attained using a pixel aggregation strategy with a 17×17 window size when mapping the spatial distribution of powdery mildew in rubber trees.

关 键 词：橡胶树白粉病 无人机 多光谱图像 遥感 堆叠集成学习模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]