基于机器学习算法的森林病虫害遥感模型对比研究  

Comparative Research on Remote Sensing Models for Forest Pests and Diseases via Machine Learning Algorithms

作　　者：郑绍鑫 何江 封成 陈积标 潘兴建 李军集[2,3] ZHENG Shaoxin;HE Jiang;FENG Cheng;CHEN Jibiao;PAN Xingjian;LI Junji(Guangxi State-owned Liuwan Forest Farm,Yulin,Guangxi 537000,China;Guangxi Forestry Research Institute/Guangxi Key Laboratory of Characteristic Non-wood Forest Cultivation&Utilization/Guangxi Engineering and Technology Research Center for Woody Spices,Nanning,Guangxi 530002,China;School of Information Engineering,China University of Geosciences(Beijing),Beijing 100083,China)

机构地区：[1]广西壮族自治区国有六万林场,广西玉林537000 [2]广西壮族自治区林业科学研究院/广西特色经济林培育与利用重点实验室/广西木本香料工程技术研究中心,广西南宁530002 [3]中国地质大学(北京)信息工程学院,北京100083

出　　处：《热带农业科学》2025年第2期80-88,共9页Chinese Journal of Tropical Agriculture

基　　金：广西自筹经费林业科技项目(No.2023GXZCLK20);广西林业科技推广示范项目(No.2024GXLK02);广西特色经济林培育与利用重点实验室自主课题(No.JA-22-03-03)。

摘　　要：森林病虫害严重威胁森林的生态功能与经济效益,如何有效识别森林病虫害并对其危害程度进行估测,对维护森林资源的健康和可持续利用至关重要。遥感技术能够实时、大范围连续动态获取植被对环境胁迫的光谱响应信息,广泛应用于森林病虫害监测。以广西六万林场森林公园为实验区,将地面调查数据及2022—2023年Sentinel-2遥感影像作为基础数据源,选取对植被颜色变化敏感的植被衰减指数(PSRI)、对植被结构变化敏感的结构不敏感色素指数(SIPI)、归一化差值红边指数(NDRE)及对植被功能变化敏感的植被光合有效辐射吸收系数(FAPAR)、归一化差值水体指数(NDWI)作为森林病虫害监测的光谱特征参数,分别利用支持向量机(SVM)和随机森林(RF)两类机器学习算法建立森林病虫害胁迫遥感监测模型,从总体精度、Kappa系数等指标对模型精度进行评价。结果表明:(1)广西六万林场森林公园的病虫害具有集聚分布、连片蔓延特点,多发生于11—12月林场森林公园北部与中部地区;(2)利用RF与SVM进行森林病虫害遥感分类的总体精度分别为63.51%92.21%、52.70%81.82%,Kappa系数分别为0.54~0.90、0.41~0.77;(3)相比SVM算法,基于RF的森林病虫害遥感监测模型精度更高,更有利于森林病虫害的监测识别。研究结果为探讨实验区的森林病虫害时空分布提供基础数据,为森林病虫害防治提供科学依据,同时为大尺度森林病虫害遥感监测提供参考。Forest pests and diseases seriously threaten the ecological function and economic benefits of forests.How to identify forest pests and diseases effectively and assessing their severity is very important for maintaining the health and sustainable use of forest resources.Remote sensing technology can continuously and dynamically obtain the spectral response information of vegetation to environmental stress in real-time and on a large scale,so it is widely used in forest pest and disease monitoring.In this study,the Guangxi State-owned Liuwan Forest Farm was taken as the experimental area,and the ground survey data and Sentinel-2 remote sensing images from 2022 to 2023 were used as the fundamental data sources.The plant senescence reflectance index(PSRI),which is sensitive to changes in vegetation function;the structure insensitive pigment index(SIPI);and the normalized difference red edge index(NDRE),which is sensitive to vegetation structure change;the fraction of absorbed photosynthetically active radiation(FAPAR);and the normalized difference water index(NDWI),which is sensitive to vegetation function changes were selected as the spectral feature parameters for monitoring forest pests and diseases.Two kinds of machine learning algorithms,support vector machine(SVM)and random forest(RF),were used to establish a remote sensing monitoring model of forest pests and disease stress,and the accuracy of the model was evaluated in terms of the overall accuracy,the Kappa coefficient,and other indicators.The results revealed that:1)the forest pests and diseases at Liuwan Forest Farm in Guangxi were characterized by an agglomeration distribution and contiguous spread,which occurred mainly in the northern and central areas of the forest farm from November to December;2)the overall accuracies of the RF and SVM remote sensing classifications of forest pests and diseases were 63.51%–92.21%and 52.70%–81.82%,respectively,and the Kappa coefficients were 0.54–0.90 and 0.41–0.77,respectively;3)compared with the SVM algorithm

关 键 词：森林病虫害 Sentinel-2影像 敏感光谱参数 随机森林 支持向量机 

分 类 号：S763[农业科学—森林保护学]