摘要
[目的]建立针对杉木人工林的单木冠幅模型,增强对杉木冠幅大小的预测准确性,并为杉木人工林的专业管理提供理论支持和依据。[方法]本研究以福建省将乐国有林场29块样地共计2005株杉木为研究对象,构建基于多种算法的杉木单木冠幅模型进行对比。从11种常用的冠幅—胸径模型中筛选出最优的逻辑斯蒂模型作为基础模型,在基础模型中添加树高、林分密度和基尼系数(Gini)等单木和林分因子作为协变量参与建模,得到广义冠幅模型,然后添加样地随机效应构建非线性混合效应模型;采用随机森林和支持向量机等随机森林算法构建杉木冠幅模型,并采用贝叶斯优化模型的超参数优化模型,得到最优超参数代入建模。采用十折交叉验证方法,以调整后的决定系数(R_(adj)^(2))和剩余均方根误差(RMSE)模型评价指标对各模型的预测效果进行评价。为进一步提高2种机器学习算法的预测精度,采用马尔可夫链蒙特卡罗算法(MCMC算法)联合随机森林与支持向量机建模,并对比基础MCMC算法与高级Stan算法在预测冠幅方面的表现。[结果]结果表明,添加样地随机效应的非线性混合效应模型预测精度高于广义模型,R_(adj)^(2)提升0.1070,RMSE下降0.0577,运用随机森林和支持向量机算法构建的冠幅模型展现出相似的预测性能,采用MCMC方法联合RF和SVM建模,MCMC链得到较好收敛,相较于基础MCMC算法,高级Stan算法各参数间自相关性显著降低,且预测精度得到提高,R_(adj)^(2)由0.6326提升至0.8495。[结论]本研究提出的基于MCMC算法和机器学习的方法有效提升了冠幅预测精度,对福建闽北现有冠幅模型的改进和应用提供了理论参考。
[Objective]The development of an individual tree crown width model for Chinese fir plantations can improve the accuracy of predicting crown width and provide scientific support and basis for the specialized management of Chinese fir plantations.[Method]This study used 29 plots with a total of 2005 Chinese fir trees from the state-owned forest farm in Jiangle County,Fujian Province.Various algorithms were employed to construct individual tree crown width models for comparison.Among 11 commonly used crown width-diameter models,the optimal logistic model was selected as the base model.Tree height,stand density,and Gini coefficient were introduced to the base model to form a generalized crown width model.A nonlinear mixed-effects model was then constructed by adding random effects of the plot.Models were also developed using random forest(RF)and support vector machine(SVM)algorithms,with the Bayesian optimization of model hyperparameters used to obtain the best hyperparameters for modeling.A ten-fold cross-validation method was employed,and model performance was evaluated using adjusted Rsquared(R_(adj)^(2))and root mean square error(RMSE).To further improve the accuracy of the two machine learning algorithms,Markov Chain Monte Carlo algorithm(MCMC)was used to combine random forest and support vector machine modeling.The performance of the base MCMC algorithm was compared with the advanced Stan algorithm in predicting crown width.[Results]The results showed that the nonlinear mixed-effects model with random effects of the plot had higher prediction accuracy than the generalized model,with R_(adj)^(2) increasing by 10.7%and RMSE decreasing by 5.77%.Crown width models constructed using RF and SVM algorithms showed similar predictive performance.When the MCMC method was used to combine RF and SVM modeling,the MCMC chain achieved good convergence.Compared to the base MCMC algorithm,the advanced Stan algorithm significantly reduced the autocorrelation between parameters,leading to improved prediction accuracy,with R_(adj)^(2) increasing from 0.6326 to 0.8495.[Conclusion]The method proposed in this study,based on the MCMC algorithm and machine learning,effectively improved the prediction accuracy of crown width.It provides valuable guidance for improving and applying existing crown width models in northern Fujian.
作者
杨自鑫
谢运鸿
孙玉军
YANG Zi-xin;XIE Yun-hong;SUN Yu-jun(Key Open Laboratory of State Forestry and Grassland Administration of Forest Resources and Environmental Management,Beijing Forestry University,Beijing 100083,China)
出处
《林业科学研究》
北大核心
2025年第4期175-186,共12页
Forest Research
基金
林业科学技术推广项目([2019]06)
国家自然科学基金项目(31870620)。
