基于XGBoost算法的森林生物量多源遥感反演  

Inversion of Forest Biomass with Multi-source Remote Sensing Data Based on XGBoost Algorithm

作　　者：蒋晋豫 王海燕 张馨之 耿佳 郭交[3] 项诗雨 JIANG Jinyu;WANG Haiyan;ZHANG Xinzhi;GENG Jia;GUO Jiao;XIANG Shiyu(Shaanxi Academy of Forestry Sciences,Xian 710082,Shaanxi,China;Shaanxi Longxiang 4D Space Information Technology Co.,Ltd.,Yangling 712100,Shaanxi,China;School of Mechanical and Electronic Engineering,Northwest A&F University,Yangling 712100,Shaanxi,China)

机构地区：[1]陕西省林业科学院,陕西西安710082 [2]陕西龙翔四维空间信息科技有限公司,陕西杨凌712100 [3]西北农林科技大学机械与电子工程学院,陕西杨凌712100

出　　处：《西北林学院学报》2025年第2期198-206,219,共10页Journal of Northwest Forestry University

基　　金：陕西省林业科学院科技创新平台建设专项(SXLK2022-0401)。

摘　　要：森林生态系统是陆地生物圈的主体,森林生物量的准确反演对于全球碳储量的研究具有重要作用。本研究以新疆西北部区域森林生物量为研究对象,使用Sentinel-1A微波遥感数据和Sentinel-2B光学遥感数据源,结合200块地面调查数据,通过提取Sentinel-2B光学遥感影像的光谱信息和植被指数,以及Sentinel-1A合成孔径雷达的后向系数,探究多源数据融合以及机器学习方法对森林生物量反演的能力。使用Relief-F算法和Pearson相关性分析进行特征优选,采用随机森林(random forest,RF)、支持向量回归(support vector regression,SVR)方法及极端梯度提升(extreme gradient boosting,XGBoost)算法,建立3种不同算法的森林生物量反演模型,并以决定系数R^(2)和均方根误差RMSE为模型评价指标,对建立的3个反演模型进行对比。结果表明:XGBoost算法表现明显优于RF和SVR算法,且在使用Pearson相关性分析进行特征优选时效果最佳,R^(2)=0.92,RMSE=11.59 t/hm^(2)。Forest ecosystems constitute the main component of the terrestrial biosphere,and accurate estimation of forest biomass is crucial for studying global carbon storage.This study took the northwestern area of Xinjiang Uyghur Autonomous Region as the research object and utilized Sentinel-1A microwave and Sentinel-2B optical remote sensing data,combined with 200 ground survey samples to explore the potential of multivariate data fusion and the capability of machine learning techniques in the inversion of forest biomass.Spectral information and vegetation indices were extracted from Sentinel-2B optical images,along with backward coefficients derived from Sentinel-1A synthetic aperture radar.Feature selection was performed using the Relief-F algorithm and Pearson correlation analysis.Different biomass estimation methods(namely Random Forest(RF),Support Vector Regression(SVR)and Extreme Gradient Boosting(XGBoost))were adopted to establish three relative inversion models of forest biomass.The determination coefficient R^(2) and root mean square error RMSE were employed as evaluation metrics to compare the performance of these models.The experimental results demonstrated that the XGBoost algorithm with feature optimization from Pearson correlation analysis outperformed RF and SVR,and R^(2) and RMSE achieved 0.92 and 11.59 t/hm^(2).

关 键 词：森林生物量 Sentinel-1A Sentinel-2B XGBoost 

分 类 号：S758.4[农业科学—森林经理学]