基于增强回归树的杉木人工林林分断面积模型研究

A Study of a Stand Basal Area Growth Model for Chinese Fir Plantations using Boosted Regression Trees

【目的】林分断面积是森林生长收获预估中的重要变量。构建基于增强回归树的林分断面积模型,并分析不同林分密度指标和不同立地指标及其交互作用对模拟的影响,为采用增强回归树模拟林分断面积提供一定的参考。【方法】研究以江西省崇义县的杉木人工林为对象,采用增强回归树构建了林分断面积模型,通过交叉验证分析了交互深度、收缩参数和回归树数量的变化对增强回归树预测的影响,并对比分析了2种立地指标(地位级指数和地位指数)、2种密度指标(每公顷株数和林分密度指数)及其交互作用对林分断面积的影响,以此筛选出林分断面积模型中最适的立地指标和密度指标。【结果】(1)经过参数优化的增强回归树能较好的模拟林分断面积,最优模型的MAB为0.5047 m^2/hm^2,RMA为2.29%,RMSE为0.9008 m^2/hm^2;(2)基于林分平均高的地位级指数对林分断面积的模拟效果优于地位指数,因此可在缺少优势高数据时采用地位级指数描述立地质量;(3)林分密度指数对林分断面积的模拟效果明显优于每公顷株数,说明包含了林木大小信息的密度指标能有效的模拟林分断面积;(4)立地和密度的交互作用能在一定程度上提升林分断面积模型精度。【结论】增强回归树能较好的模拟林分断面积,但需要对参数调优,且在需要精确预估林分断面积时,建议考虑立地和密度的交互作用。

[Objective]Stand basal area was an important variable in forest growth and yield modeling.A stand basal area model was constructed based on boosted regression trees and the effects of different stand density indicators and site indicators and the intereactions on modeling stand basal area were analyzed,which provided a reference for simulating the stand basal area using boosted regression trees.[Method]The present study focused on China fir plantations in Chongyi County,Jiangxi Province,and constructed a stand basal area model using boosted regression trees.Based on cross validation,the analyses of the effects of interaction depth,shrinkage and tree number on the prediction of boosted regression trees were performed.According to the effects of two site indicators(site class index and site index)and two density indicators(stand density index and trees per hectare)and the interactions on stand basal area were compared,the best site indicators and density indicators used in stand basal area model could be selected.[Result]First,tuned boosted regression trees simulated stand basal area well.The MAB of the best model was 0.5047 m^2/hm^2,RMA was 2.29%,and RMSE was 0.9008 m^2/hm^2.Second,the simulation of site class index based on average height of stand was better than that of site index.Thus,site quality could be described by site class index if dominant height was absent.Third,the simulation of stand density index was significantly better than that of trees per hectare,indicating that density index with tree size information could effectively simulate stand basal area.Fourth,the interaction between site and density could improved simulation to some degree.[Conclusion]Boosted regression trees could be used to model stand basal area with tuning.And the interaction between site and density should be considered if accurately estimate of stand area is needed.