机载激光雷达数据与机器学习算法的森林蓄积量估测模型构建精度评价—基于 KNN、XGBoost与RF模型反演算法

Construction Accuracy Evaluation of Forest Stock Estimation Model Based on Airborne Lidar Data and Machine Learning Algorithm

摘要:

基于激光雷达系统获取数据，旨在探索建立一个适用于湖北省的混合树种蓄积量估测模型。研究区涵盖 9 个市州及 15 个县市区 386 个样地（小班），涉及 3 种森林类型（阔叶林、针叶林和针阔混交林），划分为 5 个植被区，分别为大别山桐柏山丘陵低山、鄂西北山地丘陵、鄂东南低山丘陵、江汉平原湖泊和鄂西南山地。从点云数据中提取森林参数特征变量，结合实地调查数据，分别采用机器算法 KNN、XGBoost 和 RF 模型对森林蓄积量进行估测，采用决定系数评价模型估测精度，对估测结果进行比较分析。结果表明：（1）RF 模型的估测值与实际值较为接近，精度高于 KNN 和 XGBoost 模型。（2）不同地貌区域的森林类型估测精度存在差异，表现为针叶林估测精度高于阔叶林。估测精度与林分郁闭度、林龄、起源等因子存在相关性，林分郁闭度较高时，估测精度较高。中龄、近熟林及过熟林估测精度较高，人工林的精度高于天然林。（3）蓄积量估测值精度与实测值的区间相关，实测值趋于一定低值与高值区间时，估测精度降低。通过激光雷达数据的反演结果与地面调查数据验证，反映了模型的准确度，促进林业调查与激光雷达融合运用，需进一步比较多种模型，并探索森林分布、林木结构特征、林分因子等之间影响估测精度的相关因素。

Abstract:

Based on the data obtained by LiDAR system; this paper aims to explore and establish a model for estimating the stock of mixed tree species in Hubei Province. The study area covered 386 sample plots（small classes）in 9 cities and 17 counties and urban areas; involving three forest types（broadleaf forest; coniferous forest and coniferous mixed forest）; which were divided into five vegetation areas; namely; Tongbai Mountain hills in Dabie Mountain; mountain hills in northwest Hubei Province; mountain hills in southeast Hubei Province; lake in Jianghan Plain and mountain land in southwest Hubei Province.The forest parameter characteristic variables were extracted from the point cloud data; and combined with the field survey data; the machine algorithm KNN; XGBoost and RF models were used to estimate the forest stock; and the determination coefficient was used to evaluate the estimation accuracy of the models; and the estimation results were compared and analyzed.The results show that: 1）the estimated value of RF model is close to the actual value; and the accuracy is higher than KNN and XGBoost model. 2）The estimation accuracy of forest types in different geomorphic regions was different; and the estimation accuracy of coniferous forest was higher than that of broad-leaved forest. The estimation accuracy was correlated with stand closure; stand age; origin and other factors. The estimation accuracy was higher when stand closure was higher. The estimation accuracy of middle age; near mature forest and over mature forest was higher. The accuracy of planted forest is higher than that of natural forest. 3）The accuracy of the estimated storage volume is correlated with the interval of the true value. When the true value tends to a certain interval of low value and high value; the estimation accuracy decreases.Through the inversion results of LiDAR data and the verification of ground survey data; the accuracy of the model is reflected; and the fusion application of forestry investigation and LiDAR is promoted. It is necessary to further compare various models and explore the relevant factors affecting the estimation accuracy among forest distribution; forest structure characteristics and stand factors.