基于卷积神经网络的尾巨桉混交林胸径-树高模型  

Breast Diameter-height Models of Eucalyptus urophylla×E.grandis Mixed Plantations Based on Convolutional Neural Network

作　　者：任一平 杨梅[1] 任世奇[2] 朱慧 韦振道 伍琪[2] REN Yiping;YANG Mei;REN Shiqi;ZHU Hui;WEI Zhendao;WU Qi(College of Forestry,Guangxi University,Nanning 530000,China;Guangxi Field Scientific Observatory of Eucalyptus Forest Ecosystems,Guangxi Zhuang Autonomous Region,Nanning 530000,China)

机构地区：[1]广西大学林学院,南宁530000 [2]广西壮族自治区林业科学研究院,南宁桉树森林生态系统广西野外科学观测研究站,南宁530000

出　　处：《热带亚热带植物学报》2025年第2期140-148,共9页Journal of Tropical and Subtropical Botany

基　　金：广西科技基地和人才专项(桂科AD20325008);中央财政林业科技推广示范项目(2021TG18);广西林业科技推广示范项目(2021TG15)资助。

摘　　要：运用卷积神经网络(CNN)对尾巨桉(Eucalyptus urophylla×E. grandis)中大径材混交林的树高进行预测,为森林资源监测和评价提供理论依据。该研究以南宁桉树野外站24块尾巨桉大径材套种米老排(Mytilaria laosensis)、红锥(Castanopsis hystrix)、灰木莲(Manglietia glauca)、火力楠(Michelia macclure)形成的混交林为研究对象,结合林分优势高分树种(组),运用样本信息与卷积神经网络先验信息统计推断,经过训练得到各树种(组)适宜模型结构。使用基本一致的建模数据求解传统的树高方程,未参与建模样地作为验证集对比分析基于非线性模型或非线性混合效应6种经典模型和3种基于激活函数的CNN模型方法进行验证。结果表明,Näslund、Curtis、Logistic、Weibull、Gomperz、Korf传统模型和L-M模型(模型Ⅰ)均方根误差(RMSE)为2.5~5.6;ReLU激活函数卷积神经网络模型(模型Ⅱ)的RMSE=2.304 2,R2=0.814 9;Logistic激活函数卷积神经网络模型(模型Ⅲ)的R2=0.958 8。CNN的激活函数模型无需依赖经验模型筛选,与传统经验模型相比,基于Logistic方程的树高-胸径卷积神经网络模型决定系数高且均方根误差低,拟合精度普遍更高,能更好地拟合不同树种的生长规律,提高预测的准确性和稳定性,优化林业的生态和经济效益。The study aimed to predict the tree height of large-diameter mixed plantations of Eucalyptus urophylla×E.grandis using convolutional neural networks(CNN)and provide a theoretical basis for forest resource monitoring and evaluation.The research focused on 24 plots of mixed plantations of Eucalyptus urophylla×E.grandis intercropped with Mytilaria laosensis,Castanopsis hystrix,Manglietia glauca,and Michelia macclure at the Nanning Eucalyptus Field Station.By combining the information of dominant tree species(groups)in the stand and using statistical inference based on sample information and prior information of CNN,suitable model structures for each tree species(group)were obtained through training.Using basically consistent modeling data,traditional tree height equations were solved,and the modeling plots that did not participate in the modeling were used as a validation set for comparative analysis based on six classical models of nonlinear models or nonlinear mixed effects and three CNN models based on activation functions.The results showed that the RMSE(root-mean-square error)values of Näslund,Curtis,Logistic,Weibull,Gomperz,and Korf traditional models and the L-M model(Model I)ranged from 2.5 to 5.6.The CNN model with ReLU activation function(Model II)had a RMSE of 2.3042 and a R2 of 0.8149,while the CNN model with Logistic activation function(Model III)had a R2 of 0.9588.The activation function models of CNN do not rely on empirical model selection.Compared with traditional empirical models,the CNN model of tree height-diameter at breast height(DBH)based on the Logistic equation has a higher determination coefficient and lower root mean square error,with generally higher fitting accuracy.It can better fit the growth patterns of different tree species,improve the accuracy and stability of predictions,and optimize the ecological and economic benefits of forestry.

关 键 词：森林培育 尾巨桉 混交林 树高 卷积神经网络 

分 类 号：G63[文化科学—教育学]