河南农业科学 ›› 2025, Vol. 54 ›› Issue (2): 145-153.DOI: 10.15933/j.cnki.1004-3268.2025.02.017

• 农业信息与工程·农产品加工 • 上一篇    下一篇

基于Sentinel-2A 影像和XGBoost 模型的滇中高原地区土壤有机碳含量反演研究

严正飞1,2,杨明龙1,2,唐秀娟3,夏永华1,2,杨赈1,2,李万涛1,2   

  1. (1.昆明理工大学 国土资源工程学院,云南 昆明 650093;2.云南省高校高原山区空间信息测绘技术应用工程研究中心,云南 昆明 650093;3. 昆明市测绘研究院,云南 昆明 650091)
  • 收稿日期:2024-07-10 出版日期:2025-02-15 发布日期:2025-04-07
  • 通讯作者: 杨明龙(1982),男,贵州岑巩人,副教授,博士,主要从事3S集成应用、三维激光扫描技术、自然资源管理等研究。:S127E-mail:20130051@kust.edu.cn
  • 作者简介:严正飞(1999),男,江苏盐城人,在读硕士研究生,研究方向:遥感技术与应用。E-mail:1339673836@qq.com
  • 基金资助:
    国家自然科学基金项目(62266026)

Inversion of Soil Organic Carbon Content in the Central Yunnan Plateau Based on Sentinel⁃2A Images and XGBoost Model

YAN Zhengfei1,2,YANG Minglong1,2,TANG Xiujuan3,XIA Yonghua1,2,YANG Zhen1,2,LI Wantao1,2   

  1. (1.Faculty of Land Resources Engineering,Kunming University of Science and Technology,Kunming 650093,China;2.Yunnan Province University Spatial Information Mapping Technology Application Engineering Research Center,Kunming 650093,China;3.Kunming Institute of Surveying and Mapping,Kunming 650091,China)
  • Received:2024-07-10 Published:2025-02-15 Online:2025-04-07

摘要: 土壤有机碳(Soil organic carbon,SOC)在保持土壤肥力、促进植物生长和农业可持续发展等方面发挥着至关重要的作用,因此,高效精准地获取SOC含量非常重要。利用Sentinel-2A多光谱遥感影像数据并结合实测SOC含量、Sentinel-1后向散射系数、植被指数和地形因子数据(高程、坡度、坡向),分别使用随机森林(RF)、深度森林(DF)和XGBoost算法模型,对姚安灌区的SOC含量进行反演研究。结果表明,从不同组合的辅助变量来看,结合不同变量因子(植被指数因子、地形因子、后向散射系数因子等)有助于提高SOC含量的预测精度,尤其加入地形因子后,RF、DF和XGBoost 3种模型的R2分别提升0.052 3、0.039 8和0.068 9。从不同预测模型测算结果分析,XGBoost和DF算法模型都可以有效地进行耕地SOC含量的精准预测,其中XGBoost算法模型与M3变量组合(Sentinel-2A影像的12个波段数据、植被指数数据、Sentinel-1后向散射系数数据以及地形因子数据的组合)结合的预测能力最高[决定系数(R2)=0.810 6,均方根误差(RMSE)=1.813 2],其次是DF算法模型(R2=0.751 2,RMSE=1.925 5),而RF算法模型的预测能力相对较低(R2=0.624 5,RMSE=2.503 1)。

关键词: 土壤有机碳, Sentinel-2A, 遥感反演, 机器学习, XGBoost算法, 滇中高原

Abstract: Soil organic carbon(SOC)plays a crucial role in maintaining soil fertility,promoting plant growth,and supporting sustainable agricultural development. Therefore,efficient and accurate acquisition of SOC content is of great significance. This study utilized Sentinel⁃2A multispectral remote sensing imagery combined with measured SOC content,Sentinel⁃1 backscattering coefficients,vegetation indices and topographic factors(elevation,slope,aspect)to investigate the inversion of SOC content in the Yao’an irrigation district using Random forest(RF),Deep forest(DF),and XGBoost models.The results indicated that,from the perspective of different combinations of auxiliary variables,incorporating various factors(vegetation indices,topographic factors,backscattering coefficients,etc.)significantly improved the prediction accuracy of SOC content.Specifically,the inclusion of topographic factors increased the R2 values of the RF,DF and XGBoost models by 0.052 3,0.039 8,0.068 9,respectively.Analysis of the prediction results from different models showed that both XGBoost and DF models could effectively predict SOC content in cultivated land. Among them,the XGBoost model combined with the M3 variable set(including 12 bands of Sentinel⁃2A spectral image,vegetation indices,Sentinel⁃1 backscattering coefficients,and topographic factors) achieved the highest prediction accuracy(R2=0.810 6,RMSE=1.813 2),followed by the DF model(R2=0.751 2,RMSE=1.925 5),while the RF model exhibited relatively lower predictive performance(R2=0.624 5,RMSE=2.503 1)

Key words: Soil organic carbon, Sentinel?2A, Remote sensing inversion, Machine learning, XGBoost algorithm, Central Yunnan Plateau

中图分类号: