摘要
2021年7—8月份,在陕西省延安市黄陵县双龙林场,以栎类次生林为研究对象,设置21块20 m×30 m的固定样地,调查样地内胸径5 cm及以上所有乔木树种种名、胸径、树高、株数、冠幅、第一活枝高等因子以及样地内每株乔木的坐标等数据;依据对21块样地调查的数据,以单木地上碳储量为评价指标,以林分因子、单木因子、立地因子为影响因素,分别构建普通最小二乘模型(OLS)、地理加权回归模型(GWR)、地理加权随机森林模型(GRF)、地理加权神经网络模型(GWANN),对研究区的单木地上碳储量进行模拟估测;采用决定系数、均方根误差、平均绝对误差对4种模型的拟合效果进行评价,利用全局莫兰指数检验4种模型残差的空间自相关性,分析地理加权机器学习模型在处理林木空间非线性关系方面的应用。结果表明:使用地理加权的模型,拟合效果明显优于普通最小二乘模型,其中地理加权神经网络模型的拟合效果最好(决定系数为0.980、均方根误差为0.169,平均绝对误差为0.119)。普通最小二乘模型和地理加权回归模型的残差,表现出明显的自相关性;地理加权随机森林模型和地理加权神经网络模型的残差,则表现为随机分布。3种使用地理加权的模型(地理加权回归模型、地理加权随机森林模型、地理加权神经网络模型)大幅度降低了模型的残差自相关性,减少了残差空间上聚集分布情况的出现。综合分析研究结果,与地理加权回归模型相比,地理加权随机森林模型和地理加权神经网络模型2种地理加权机器学习模型,能够更好地反映单木地上碳储量与各影响因素间的复杂关系,预测单木地上碳储量的效果更好,具有良好的应用潜力。
In July and August 2021,a study was conducted in Shuanglong Forest Farm,Huangling County,Yan’an City,Shaanxi Province,focusing on Quercus secondary forests.Twenty-one fixed plots measuring 20 m×30 m were established to survey the species,diameter at breast height(DBH),tree height,tree count,crown width,first live branch height,and coordinates of all trees with a DBH of 5 cm and above in the plots.Based on the data collected from the 21 plots,individual tree carbon stock was used as the evaluation indicator,with stand-level factors,individual tree factors,and site factors as influencing variables.Ordinary Least Squares(OLS),Geographic Weighted Regression(GWR),Geographic Weighted Random Forest(GRF),and Geographic Weighted Artificial Neural Network(GWANN)models were separately constructed to simulate and estimate the carbon stock of individual trees in the study area.The fitting performance of the models was evaluated using coefficients of determination,root mean square error,and mean absolute error.Global Moran’s Index was used to test the spatial autocorrelation of the residuals of the four models,and the application of geographic weighted machine learning models in handling spatial non-linear relationships among forest trees was analyzed.The results showed that the fit of the models improved significantly when using the geographic weighted approach compared to the OLS model,with the GWANN model exhibiting the best fit(determination coefficient of 0.980,root mean square error of 0.169,and mean absolute error of 0.119).Residuals of the OLS and GWR models showed significant autocorrelation,while residuals of the GRF and GWANN models exhibited a random distribution.The three models using geographic weighting(GWR,GRF,GWANN)substantially reduced the autocorrelation of the residuals and decrease the spatial clustering of residuals.Overall,the GRF and GWANN models outperformed the GWR model in reflecting the complex relationships between above-ground carbon stock on individual trees and influencing factors,with better predictive accuracy and promising application potential.
作者
魏江涛
卜元坤
周建云
李卫忠
王明杰
Wei Jiangtao;Bu Yuankun;Zhou Jianyun;Li Weizhong;Wang Mingjie(North West Agriculture and Forestry University,Yangling 712100,P.R.China;Yan’an Qiaoshan State Forest Administration Shuanglong State Ecological Experimental Forest Farm)
出处
《东北林业大学学报》
CAS
CSCD
北大核心
2024年第6期98-105,共8页
Journal of Northeast Forestry University
基金
国家自然科学基金青年项目(32001311)
陕西省林业科技创新计划专项(SXLK2021-0208)。
关键词
栎类次生林
单木碳储量
地理加权回归
Quercus secondary forests
Individual tree carbon stock
Geographic weighted regression
