基于Landsat 8数据的人工林地上生物量估测模型研究  

Estimation Model of Aboveground Biomass in Plantations Based on Landsat 8 Data

作　　者：闫宇 邓焯 李斌 赵天忠[1,2,3] YAN Yu;DENG Zhuo;LI Bin;ZHAO Tianzhong(College of Information,Beijing Forestry University,Beijing 100083,China;Engineering Research Center for Forestry-Oriented Intelligent Information Processing of National Forestry and Grassland Administration,Beijing 100083,China;Institute of Forestry Informatization,Beijing Forestry University,Beijing 100083,China;Institute of Forestry and Pomology,Beijing Academy of Agriculture and Forestry Sciences,Beijing 100093,China;Beijing Yanshan Forest Ecosystem Research Station,National Forest and Grassland Administration,Beijing 100093,China)

机构地区：[1]北京林业大学信息学院,北京100083 [2]国家林业草原林业智能信息处理工程技术研究中心,北京100083 [3]北京林业大学林业信息化研究所,北京100083 [4]北京市农林科学院林业果树研究所,北京100093 [5]国家林业和草原局北京燕山森林生态系统国家定位观测研究站,北京100093

出　　处：《西北林学院学报》2024年第5期53-60,77,共9页Journal of Northwest Forestry University

基　　金：国家重点研发计划项目(2017YFD0600906)。

摘　　要：为提高森林资源管理的效率和精度,探讨特征变量选择与新型机器学习算法结合建立桉树人工林地上生物量估测模型的精度。以广西高峰林场为研究区,以Landsat 8遥感数据结合实测样地数据,使用Pearson相关性分析法结合随机森林的特征变量选择方法,分别构建基于多元线性回归(MLR)、K最邻近(KNN)、随机森林(RF)和极端梯度提升(XGBoost)算法的森林地上生物量估测模型,使用模型评价指标对比不同模型的精度。结果表明,XGBoost模型拟合精度最高,验证结果R^(2)为0.75、RMSE为30.15 t/hm^(2)、MAE为20.27 t/hm^(2);RF、KNN和MLR模型次之,R^(2)分别为0.69、0.54和0.52。利用Pearson相关性分析法结合随机森林相较于仅使用随机森林筛选变量的方法,R^(2)提高了27.12%、RMSE降低了11.44 t/hm^(2)、MAE降低了8.70 t/hm^(2)。采用机器学习方法的模型比多元线性回归模型更有优势,其中新型机器学习算法XGBoost在生物量估测方面有巨大潜力。Pearson相关性分析结合随机森林的特征选择方法能够减少冗余变量对模型估测精度的影响,有效提高模型预测性能。In order to improve the efficiency and accuracy of forest resource management,the accuracy of the aboveground biomass estimation model of Eucalyptus plantation was explored by combining feature variable selection with new machine learning algorithm.Taking Guangxi Gaofeng Forest Farm as the study area,the aboveground biomass estimation model based on multiple linear regression(MLR),K-nearest neighbor(KNN),random forest(RF)and extreme gradient boosting(XGBoost)algorithms were constructed by using the Pearson correlation analysis method in combination with the actual sample data from Landsat 8 remote sensing data and the feature variable selection method of RF.The accuracies between different methods were compared using the model evaluation index.The results showed that the XGBoost model had the highest fitting accuracy,with R^(2) of 0.75,RMSE of 30.15 t/hm^(2),and MAE of 20.27 t/hm^(2);the RF,KNN,and MLR models followed with R^(2) of 0.69,0.54,and 0.52,respectively.When using Pearson correlation analysis method combined with RF,R^(2) increased by 0.16,RMSE decreased by 11.44 t/hm^(2) and MAE decreased by 8.70 t/hm^(2) compared with the method using RF only.Models using machine learning methods were more advantageous than multivariate linear regression model.The new machine learning algorithm XGBoost has great potential in biomass estimation.Pearsons feature selection method combined with random forest can reduce the influence of redundant variables on the estimation accuracy of the model and effectively improve the prediction performance of the model.

关 键 词：森林地上生物量 变量筛选 机器学习 XGBoost Landsat 8 

分 类 号：S758.51[农业科学—森林经理学]