为探究不同模型对林分平均胸径的预测精度,使用贵州省桂花国有林场马厂工区同步获取的机载激光雷达点云数据和地面实测样地数据,通过提取样地水平的点云特征变量,采用方差膨胀因子分析和皮尔逊相关性检验进行自变量选择,建立机器学习模型估测样地平均胸径。结果表明:1)点云特征变量,如平均冠层高度和高度偏态与林分平均胸径有很强的相关性。2)机器学习模型(随机森林、支持向量机、最近邻算法)优于多元线性回归模型,其中,随机森林的拟合效果最好。随机森林的决定系数(R²)为0.71,均方根误差(RMSE)为2.50。3)通过柳杉纯林、针叶混交林、针阔混交林、马尾松纯林4种森林类型的林分平均胸径预测值与实际值差值,进一步证实随机森林模型精度最高,拟合效果最好。利用机载激光雷达点云数据提取点云特征变量,并构建基于机器学习算法的林分平均胸径估测模型是可行的,该方法的精度能满足森林资源调查的应用需求,可作为辅助林业调查工作的技术手段。

In order to explore the prediction accuracy of different models on the average diameter at breast height of forest stands,airborne LiDAR point cloud data and ground measured sample plot datawere obtained simultaneously by the Machang working area of Guihua State-owned Forest Farm in Guizhou Province.By extracting point cloud feature variables at the sample plot level,a machine learning model is used toestimate the average diameter at breast height of the sample plot,variance inflation factor analysis and Pearson correlation test are used to select independent variables.The results indicate that:1)Point cloud feature variables show a strong correlation with the average diameter at breast height of the forest stand,such as the average canopy height and height skewness.2)Machine learning models(random forest,support vector machine,nearest neighbor algorithm)outperform multiple linear regression models,with random forest having the best fitting performance.The determination coefficient(R²) for the random forest model is 0.71,withthe root mean square error(RMSE)of 2.50.3)The difference between the predicted and actual average diameter at breast height of four forest types:Cryptomeria forest,mixed coniferous forest,mixed coniferous and broad-leaved forest,and Pinusmassoniana forest further confirms that the random forest model has the highest accuracy and the best fitting effect.In summary,it is feasible to extract point cloud feature variables using airborne LiDAR point cloud data and construct a forest average diameter estimation model based on machine learning algorithms.The accuracy of this method meets the application requirements of forest resource investigation and can be used as a technical means to assist in forestry investigation work.