Stacking集成学习模型在森林蓄积量估测中的应用  

Application of Stacking Ensemble Learning Model in Forest Volume Estimation

作　　者：邹泽林 程霞 刘紫薇 黄鑫 Zou Zelin;Cheng Xia;Liu Ziwei;Huang Xin(Central South Investigation and Planning Institute of National Forestry and Grassland Administration,Changsha,410014)

机构地区：[1]国家林业和草原局中南调查规划院,长沙410014

出　　处：《湖北林业科技》2023年第4期52-57,共6页Hubei Forestry Science and Technology

摘　　要：为了探索Sentinel-2遥感数据在估测森林蓄积量中的适用性,以及开发一种提高蓄积量估测精度的集成学习算法,选择江西省兴国县为研究区,以Sentinel-2为遥感数据源,利用Boruta算法进行特征选择后开发了一种Stacking集成学习模型,并且与MLR、KNN、SVR和RF四种基础模型进行对比。结果表明,相比于MLR模型,机器学习模型具有更强的蓄积量估测能力,利用机器学习模型估计森林蓄积量的RMSE降低了18.02~22.50 m^(3)·hm^(-2),rRMSE降低了9.01%~11.25%。并且相比于基础模型,利用Stacking算法将4种模型进行集成后,模型的均方根误差进一步降低了11.95~7.47 m^(3)·hm^(-2),说明Stacking集成学习算法可以有效的提高森林蓄积量的估测性能。In order to explore the applicability of Sentinel-2 remote sensing data in estimating forest stock and to develop an integrated learning algorithm to improve the accuracy of stock estimation,Xingguo County,Jiangxi Province was selected study area,uses Sentinel-2 as the remote sensing data source,develops a stacking integrated learning model using Boruta algorithm for feature selection,and compares it with four basic models,namely MLR,KNN,SVR and RF.The results showed that compared with the MLR,KNN,SVR and RF models,the stacking integrated learning model was more efficient than the MLR model.The machine learning model has stronger accumulation estimation ability than the MLR model,and the RMSE of forest accumulation estimation using the machine learning model was reduced by 18.02~22.50 m^(3)·hm^(-2)and the rRMSE was reduced by 9.01~11.25 percentage points.In addition,the RMSE of the model was further reduced by 11.95~7.47 m^(3)·hm^(-2)after integrating the four models using the Stacking algorithm compared with the base model,indicating that the Stacking integrated learning algorithm can effectively improve the estimation performance of forest stock.

关 键 词：森林蓄积量 Sentinel-2 集成学习 遥感估测 

分 类 号：S771.8[农业科学—森林工程]