Research on information extraction of forest fire damage based on multispectral UAV and machine learning
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>mSRrededge>NDVIrededge>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.
NETL
NSTL
万方数据
