基于集成学习和Sentinel-2的落叶松毛虫虫害区识别

Dendrolimus Superans Infected Area Identification Based on Ensemble Learning Model and Sentinel-2 Data

为实现快速、高效地监测落叶松毛虫(Dendrolimus superans)虫害爆发状况,以黑龙江省乡南经营所林场8林班为研究区,以2018年Sentinel-2遥感影像为数据源,对该林班的落叶松毛虫虫害区进行识别。提取预处理后影像的原始光谱特征(8个)、光谱指数特征(12个)与纹理特征(8个),基于方差分析(Analysis of Variance,ANOVA)与极度梯度提升(eXtreme Gradient Boosting,XGBoost)分类器对上述特征降维并按重要性排序,通过集成学习分类算法(随机森林分类器和XGBoost分类器)进行虫害区识别和精度比较。研究结果表明,(1)应用重要性前14位特征的XGBoost模型对虫害区识别的表现最为理想,总体分类精度为95%(Kappa系数为86%),高于随机森林分类器的93%(应用重要性前10的特征);(2)重要性前14的特征名称由大到小为EVI1、Mean、MTCI、GNDVI、Variance、B4、B2、Homogeneity、B3、CRI1、EVI2、B8、B5和CRE。研究结果可实现落叶松毛虫虫害区的高效识别,为东北林区的虫害防治决策制定提供依据。

In order to quickly and efficiently monitor the outbreaks of Dendrolimus superans,the 8th forest compartment of South Management Office Forest Farm in Heilongjiang Province was taken as the research area,while the Sentinel-2 remote sensing image in 2018 was used as data source to identify the Dendrolimus superans infestation area in the forest compartment.The original spectral features(8),spectral index features(12)and the texture features(8)were extracted from the preprocessed image.Based on ANOVA and XGBoost classifiers,all features were dimensionally reduced and sorted by importance.The ensemble learning classification algorithm(Random Forest classifier and XGBoost classifier)was used to identify pest areas and compare their accuracy.The results showed that:(1)the XGBoost model with the top 14 important features was the most ideal for the identification of pest areas,and the overall accuracy reached to 95%(Kappa coefficient=86%),which were higher than the 93%of Random Forest(the top 10 features in order of importance);(2)the top 14 feature names were:EVI1,Mean,MTCI,GNDVI,Variance,B4,B2,Homogeneity,B3,CRI1,EVI2,B8,B5 and CRE.This method can achieve efficient identification of Dendrolimus superans infestation areas,which can provide a basis for decision-making on pest control in northeast forest.