依据BP神经网络的机载LiDAR数据估算林分平均高  

Estimation of Mean Forest Stand Height from Airborne LiDAR Data Using BP Neural Network

作　　者：周蓉 赵天忠[1] 吴发云 Zhou Rong;Zhao Tianzhong;Wu Fayun(Beijing Forestry University,Beijing 100083,P.R.China;Academy of Forest Inventory and Planning of State Forestry Administration)

机构地区：[1]北京林业大学信息学院,北京100083 [2]国家林业和草原局调查规划设计院

出　　处：《东北林业大学学报》2021年第9期60-66,共7页Journal of Northeast Forestry University

基　　金：国家林业和草原局2020年行业管理专项业务经费(2020-21-89)。

摘　　要：以东北虎豹国家公园范围内的针叶纯林为研究对象,结合2018年9月机载LiDAR点云数据和同步地面调查数据,提取半径为15 m的圆形采样尺度下的LiDAR点云特征变量为数据基础,采用BP神经网络算法、逐步回归法分别构建林分算术平均高模型和林分加权平均高模型,实现对林分平均高的估测。其中在利用BP神经网络算法构建模型时分别选择了贝叶斯正则化算法和L-M算法作为神经网络训练算法。结果表明:BP神经网络算法对数据具有更好地解释能力,其构建的林分平均高模型相关系数(R^(2))均在87%以上,高于逐步回归法构建的林分平均树高模型;林分加权平均高模型精度更高,用样地加权平均高作为实测值时,采用逐步回归算法、BP神经网络L-M算法、BP神经网络贝叶斯正则化算法构建的模型的检验样地数据的决定系数(R^(2))分别为0.858、0.919、0.908,树高估测精度(P)分别为88.6%、89.8%、91.2%,与以林分算术平均高作为实测值构建的估测模型相比,决定系数(R^(2))分别提升了4.9%、3.7%、3.4%,估测精度(P)分别提升了2.9%、2.4%、1.5%;BP神经网络的不同训练函数之间无明显性能差异,两种不同训练方法构建的林分平均高模型的决定系数R^(2)及树高估测精度(P)略有差异,但整体相差较小。By taking the pure coniferous forest in the Northeast Tiger and Leopard National Park as the research object,combining the airborne LiDAR point cloud data obtained in September 2018 and the synchronized ground survey data,andextracting the LiDAR point cloud features at a circular sampling scale with a radius of 15 m as the data,the BP neural network algorithm and the stepwise regression method were used to construct the stand arithmetic average height model and the stand weighted average height model to realize the estimation of the average stand height.Among them,the Bayesian regularization algorithm and the L-M algorithm were selected as the neural network training algorithm when using the BP neural network algorithm to build the models.The results show that the BP neural network algorithm has a better ability to explain the data,and the correlation coefficient R^(2) of the average stand height model constructed by it is above 87%,which is higher than the average stand height model constructed by the stepwise regression method.The weighted average high model has higher accuracy.When the weighted average high of the sample plot is used as the measured value,the test set of the model constructed by the stepwise regression algorithm,the BP neural network LM algorithm,the BP neural network Bayesian regularization algorithm andthe coefficients of determination(R^(2))of the sample plot data are 0.858,0.919 and 0.908,and the tree height estimation accuracy P are 88.6%,89.8%and 91.2%,respectively.Compared with the estimation model constructed with the stand arithmetic average height as the measured value,the coefficients of determination(R^(2))are increased by 4.9%,3.7%and 3.4%,respectively,and the estimation accuracies(P)are increased 2.9%,2.4%and 1.5%respectively.There was no significant performance difference between the different training functions of the BP neural network.The determination coefficient R^(2) and the tree height estimation accuracy P of the constructed forest average height model constructed by two

关 键 词：激光雷达 点云数据 林分平均高 BP神经网络 

分 类 号：S758[农业科学—森林经理学]