基于Sentinel-2卫星影像和土壤变量的盐渍化土壤水溶性盐基离子含量反演  

Inversion of Soil Water-Soluble Salt Ion Content in Saline-Alkali Soil Based on Sentinel-2 Satellite Imagery and Soil Variables

作　　者：谭旺 刘义 董建华 杨阳 黄介生[1] 敖畅 曾文治 TAN Wang;LIU Yi;DONG Jian-hua;YANG Yang;HUANG Jie-sheng;AO Chang;ZENG Wen-zhi(School of Water Resources and Hydropower Engineering,Wuhan University,Wuhan 430072,Hubei Province,China;College of Agricultural Science and Engineering,Hohai University,211100,Jiangsu Province,China)

机构地区：[1]武汉大学水利水电学院,湖北武汉430072 [2]河海大学农业科学与工程学院,江苏南京211100

出　　处：《中国农村水利水电》2024年第7期210-217,228,共9页China Rural Water and Hydropower

基　　金：国家重点研发计划(2021YFD1900805)。

摘　　要：本研究旨在探讨结合多光谱遥感技术和土壤物理化学属性进行土壤水溶性离子含量预测的可行性。研究区域位于新疆南部的盐渍化土壤区域,测定土壤中主要水溶性盐基离子含量(K^(+)、Na^(+)、Ca^(2+)、Mg^(2+)、HCO_(3)^(-)、Cl^(-)、SO_(4)^(2-)),并应用随机森林(Random Forest,RF),梯度提升回归树(Gradient Boosting Regression,GBR)和极端梯度提升(Extreme gradient boosting,XGBoost)3种机器学习算法构建基于遥感光谱特征及土壤信息的土壤离子含量反演模型,同时对比了纳入与未纳入土壤变量的模型预测精度。结果表明:仅使用多光谱遥感数据作为输入变量时,3种模型均仅能区分土壤离子含量的高低水平,但对各离子含量进行精确预测的能力有限。将土壤变量纳入模型后,预测精度均得到显著提升。在所选用的3种方法中,随机森林模型的预测精度最高,XGBoost次之,GBR模型的精度最低。就各个离子的预测而言,Mg^(2+)、Ca^(2+)和Na^(+)含量的预测精度较高且模型表现较为稳定;SO_(4)^(2-)、Cl^(-)和K^(+)的预测表现一般,具备定量预测能力;而HCO_(3)^(-)含量的预测仅GBR模型具有一定程度的可行性。不同离子的最优预测模型存在差异,其中随机森林模型对K^(+)、Mg^(2+)和Cl^(-)三种离子的反演效果最佳;XGBoost模型在Ca^(2+)、Na^(+)和SO_(4)^(2-)三种离子的反演中表现为最优;GBR模型则在HCO_(3)^(-)离子反演中展现了较好的性能。特别地,Mg^(2+)、Ca^(2+)和Na^(+)含量的最优相对分析误差分别为2.829、1.951和1.870,说明这些模型对于这3种离子含量的预测具有较高的可靠性。研究成果可以为干旱区土壤盐分主要离子含量的区域尺度预测提供科学参考。This study aims to explore the feasibility of estimating the content of water-soluble soil ions by combining multispectral remote sensing technology with soil physicochemical properties.The research area is situated in the saline soil regions of southern Xinjiang,where the concentrations of major water-soluble cations and anions(K^(+)、Na^(+)、Ca2^(+)、Mg^(2+)、HCO_(3)^(-)、Cl^(-)、SO_(4)^(2-))were measured.Machine learning algo⁃rithms such as Random Forest (RF), Gradient Boosting Regression (GBR), and Extreme Gradient Boosting (XGBoost) were employed toconstruct soil ion content inversion models based on remote sensing spectral features and soil information. Additionally, the study comparedthe estimation accuracy of models incorporating soil variables with those that did not. Results indicate that when only multispectral remotesensing data were used as input variables, all three models could only differentiate between high and low levels of soil ion content, with limit⁃ed ability to accurately estimate the concentrations of individual ions. Incorporation of soil variables into the models significantly enhanced es⁃timation accuracy. Among the methods used, the RF model exhibited the highest prediction accuracy, followed by XGBoost, and GBR hadthe lowest accuracy. Regarding the estimation of specific ions, the concentrations of Mg^(2+), Ca2^(+), and Na^(+) were predicted with relatively highprecision and model performance was stable;SO_(4)^(2-) , Cl^(-) , and K^(+) showed moderate performance with quantitative prediction capabilities;whereas HCO_(3)^(-) content estimation was only feasible to a certain extent with the GBR model. Optimal models varied for different ions, with theRF model providing the best inversion results for K^(+), Mg^(2+), and Cl^(-);the XGBoost model excelling in the inversion of Ca2^(+), Na^(+), and SO_(4)^(2-);and the GBR model performing well for HCO_(3)^(-) inversion. Notably, the optimal relative analysis errors for Mg^(2+), Ca2^(+), and Na^(+) content estima⁃tion were 2.829

关 键 词：土壤 盐渍化 盐基离子 Sentinel-2 机器学习 反演模型 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] S156.41[自动化与计算机技术—控制科学与工程]