ABSTRACT

In forest resource inventory, tree height is often estimated by easily measurable diameter from height-diameter model. In this study, we tried to use random forest (RF), an important machine learning method, to model individual-tree height. Results showed that the optimized RF model had better fitting and prediction accuracy (R²= 0.8146 and RMSE = 2.2527 m). In terms of relative importance, diameter at breast height (DBH) was the most important factor, followed by neighborhood-related variables and other variables related to environmental conditions. Further, tree height was generally positively affected by DBH, mean diameter of neighbors, DBH dominance, number of neighbors, and mean annual precipitation, but negatively affected by elevation. The results indicated that the RF-based height model was statistically reliable and highly accurate, and it had strong interpretability with ecological significance. Our study will provide a new perspective for the application of machine learning algorithms to forest dynamic modeling.

摘要

在森林资源清查中，树木高度通常通过高度-直径模型中易于测量的直径来估计。在本研究中，我们尝试使用随机森林（RF）（一种重要的机器学习方法）来建模单棵树的高度。结果表明，优化后的RF模型具有更好的拟合和预测精度（R ² = 0.8146和RMSE） = 2.2527 m）。就相对重要性而言，胸径（DBH）是最重要的因素，其次是邻里相关变量和其他与环境条件相关的变量。此外，树高通常受胸径、相邻树的平均直径、胸径优势、相邻树的数量和年平均降水量的正向影响，但受海拔的负向影响。结果表明，基于RF的高度模型统计可靠、准确度高，可解释性强，具有生态意义。我们的研究将为机器学习算法在森林动态建模中的应用提供新的视角。