基于多光谱无人机及机器学习的林木火灾受损信息提取研究

Research on information extraction of forest fire damage based on multispectral UAV and machine learning

为探究中小尺度森林火灾过火区域林木受损程度的准确提取,以2020年5月13日云南省安宁市青龙街道森林火场为研究对象,通过精灵4多光谱无人机获取火场影像,借助红边及近红外波段构建植被指数,结合纹理指标建立影像特征参数,利用机器学习中常用的随机森林(random forest,RF)和支持向量机(support vector machine,SVM)方法提取烧毁、烧死、烧伤及未伤林木空间分布信息,并探讨2种方法对于多光谱无人机遥感林木受损信息提取的精度。结果表明:不同受损程度的林木在红边波段和近红外波段范围内反射率差异较大,但以此构建的植被指数分离能力不同,呈现NDVI>mSR rededge>NDVI rededge>PSRI。基于影像光谱及纹理等多特征的林木受损程度提取方法中,RF精度明显优于SVM,总精度达89.76%,Kappa系数为0.85,相比SVM分别提升4.41%和6.25%。多光谱无人机可用于小范围典型森林火灾区域林木受损程度信息精确提取,而对于大面积范围的林木火灾受损信息的精确提取,综合多光谱无人机数据及多光谱卫星影像数据是解决问题的方向。

To investigate the accurate extraction of forest tree damage in small-and medium-scale forest fire areas,the forest fire site in Qinglong Street,Anning City,Yunnan Province,was selected as the research object.The fire site images were acquired by a Genie 4 multispectral unmanned aerial vehicle(UAV)on May 13,2020.The vegetation indices were constructed with the help of red-edge and near-infrared bands,and image feature parameters were established by combining texture indicators.The random forest(RF)and support vector machine(SVM)methods,which are commonly used in machine learning,were used to extract the spatial distribution information of burnt,dead,damaged and unburned trees,and to explore the accuracy of the two methods for extracting the damage information of remote sensing trees by multispectral UAV.Results are as follows.There were great differences of the reflectance of forest trees with different damage levels in the red-edge and near-infrared bands.Additionally,the separation ability of vegetation indices constructed by this method was different,showing NDVI>mSR rededge>NDVI rededge>PSRI.Among the methods of extracting forest tree damage levels based on multiple features such as image spectra and textures,RF accuracy was significantly better than that of SVM,with a total accuracy of 89.76%and Kappa coefficient of 0.85,which were 4.41%and 6.25%higher than those of SVM,respectively.Multispectral UAV can be used for accurate extraction of forest damage information in small-scale typical forest fire areas,whereas for large-area regions,multispectral UAV data can be the prospective solution.