基于多光谱卫星影像与机器学习算法的松材线虫病受害林分识别研究

松材线虫病(Pine Wilt Disease, PWD)被称为“松树癌症”,具有高传染率和高死亡率，对我国森林资源构成了严重的威胁，对我国的经济、社会和生态造成了重大损失。及时发现并清理疫木是遏制松材线虫病蔓延的有效手段，精准监测疫木是防控松材线虫病的前提，但是现阶段缺少大面积识别松材线虫病疫木的技术方法。本文旨在探索哨兵-2号与Landsat-8遥感卫星影像对受害松林的识别能力，采用随机森林(Random Forest, RF)、支持向量机(Support Vector Machine, SVM)、决策树(Decision Tree, DT)和极端梯度提升(Extreme Gradient Boosting, XGBoost)等4种机器学习算法建立了松材线虫病监测模型。结果表明：基于哨兵-2号影像数据建立的监测模型对受害松林的识别准确率高于Landsat-8遥感卫星影像，其中基于10 m分辨率的影像数据建立的监测模型识别准确率最高，随机森林、决策树、支持向量机和极端梯度提升等算法建立模型的准确率分别达到了79.3%、76.2%、78.7%和78.9%。在3种不同的影像数据集中，RF...

Based on multispectral satellite images and machine learning algorithms to recognize pine wood nematode disease affected stands

Pine Wilt Disease (PWD), known as the cancer of pine, has a high infection rate and high mortality rate, posing a serious threat to China''s forest resources and causing significant economic, social and ecological losses to China. Timely detection and cleaning of infected wood is an effective means to curb the spread of PWD, and accurate monitoring of infected wood is a prerequisite for prevention and control of PWD. However, there is a lack of technical methods to identify infected wood of PWD in a large area at this stage. This paper aims to explore the recognition ability of Sentinel-2 and Landsat-8 on infected wood of PWD and to establish PWD monitoring models using four machine learning algorithms, including Random Forest (RF), Support Vector Machine (SVM), Decision Tree (DT) and Extreme Gradient Boosting (XGBoost). The results showed that the monitoring models based on Sentinel-2 image had higher recognition accuracy than Landsat-8 image. In addition, the models with highest recognition accuracy were based on 10 m resolution image, and these models built by RF, DT, SVM and XGBoost reached accuracy at 79.3%, 76.2 %, 78.7% and 78.9%, respectively. The accuracy, kappa coefficient and ROC values of the RF, SVM and XGBoost were close and all significantly better than DT in the three different images. Green band, red band, short-wave NIR band and long-wave NIR band in the spectral features and NBRI, NGRDI, TVI, NDVI and PSSR in the vegetation indices had the highest contribution values to PWD surveillance models. Mean Decrease in Impurity (MDI) was the most effective method in filtering the feature parameters, and the number of features was reduced from 50 to 35. PWD surveillance models established in this paper provide technical support for scientific prevention and control of PWD.