摘要
土壤有机碳(SOC)是陆地生态系统碳循环的重要组成部分,也是评价区域土壤质量、土地退化程度和作物产量的重要指标。高寒生态系统土壤有机碳含量估算,对于高寒地区土壤碳库核算和土壤质量评价等都具有重要意义。本研究以青藏高原三江源区作为研究区,基于野外采集的272个土壤样本的SOC和土壤光谱室内测试数据,首先对原始光谱数据进行一阶微分(FD)、二阶微分(SD)、倒数对数(RL)、去包络线(CR)和多元散射校正(MSC)等多种数学变换;然后基于8种光谱变换数据与SOC含量的相关性分析提取特征波段,利用多元线性回归(MLR)、偏最小二乘回归(PLSR)、支持向量机(SVM)和随机森林(RF)4种方法,分别构建SOC含量的高光谱反演模型;对各种模型的模拟精度和稳定性进行评价,明确SOC含量反演的最佳光谱变换和模型组合模式。结果表明:三江源区SOC含量较高,且不同植被类型和不同土壤类型的SOC含量差异较大;总体上,单一数学变换形式中基于一阶微分(FD)变换构建的反演模型的模拟精度最高(尤其是FD-RF组合模型,其验证集R^(2)=0.86,RMSE=8.40,RPD=2.64);多种数学变换组合相较于单一的数学变换,其模拟精度得到进一步提高(如CRFD-RF组合,R^(2)=0.87,RMSE=8.03,RPD=2.76);4种模拟模型中,随机森林总体模拟精度最高,并且CRFD-RF组合模型的模拟精度最高,MSCFD-RF次之。本研究可为利用高光谱遥感进行高寒地区土壤有机碳含量反演和动态监测提供理论依据。
Soil organic carbon(SOC)is an important part of the carbon cycle of the terrestrial ecosystem,and also an important indicator to evaluate regional soil quality,land degradation and crop yield.The estimation of soil organic carbon content in alpine ecosystem is of great significance for soil carbon pool accounting and soil quality evaluation in alpine regions.Visible and near infrared reflectance spectroscopy(Vis-NIRS)has been proven to be an efficient method for predicting soil properties,and the combination of spectral transformation and models can improve the simulation accuracy of SOC,but the best combination of mathematical transformation methods and inversion model are unknown for alpine ecosystem soil.In this study,Three Rivers Source Region(TRSR)of Qinghai-Tibet Plateau as the study area and 272 soil field samples were collected.The collected samples air-dried,sieved through 0.25 mm mesh,handpicked to remove roots.Then the well-done soil samples were used to indoor chemical analysis and soil spectrum test.SOM was determined by potassium dichromate sulfuric acid external heating in laboratory.The soil spectral curves were tested using US ASD FieldSpec 4 Standard Res in a dark room.The original spectral curve undergoes a series of mathematical transformations,such as first-order differential(FD),second-order differential(SD),reciprocal logarithm(RL),Continuum removal(CR),multivariate scattering correction(MSC),etc.The characteristic band were selected based on correlation analysis between SOC content and spectral reflectance.The hyperspectral inversion models of SOC content were built using four methods,namely,multiple linear regression(MLR),partial least squares regression(PLSR),support vector machine(SVM)and random forest(RF).Meanwhile,the best spectral transformation and model combination mode for SOC content inversion were determined by the comparison of all the models based on the validation sets’simulation accuracy and stability evaluation.Results showed that:(1)SOC content in the Three Rivers Source Region was high and the average value was 34.68 g·kg^(-1),which was significantly higher than the first grade of national standard for nutrient grade I(23.2 g·kg^(-1)g)of China.Meanwhile,SOC content in different vegetation types or different soil types was various different,and the order of soil organic carbon content from large to small in different vegetation types is Alpine meadow,forest,Alpine meadow grassland,Alpine grassland,temperate grassland,farmland and desert grassland.SOC content of Alpine meadow(average value was 48.59 g·kg^(-1))was significantly larger than desert grassland(15.38 g·kg^(-1)),statistical analysis in this study also found the coverage and biomass of Alpine meadow(90%,130 g·kg^(-1))significantly larger than desert grassland(46%,84 g·kg^(-1)).SOC content of Meadow swamp soil was highest with 59.12 g·kg^(-1),grass felt soil also was high with 50.83 g·kg^(-1),the content of SOC in chestnut soil and calcic soil is lowest with 21.48 g·kg^(-1),and 20.77 g·kg^(-1),respectively.(2)In general,the inversion model constructed by first-order differential(FD)transformation obtained the highest simulation accuracy among the single mathematical transformation form(especially the model combination of FD-RF,with R^(2)=0.86,RMSE=8.40,RPD=2.64;in its validation set).Because FD transform can significantly improve the characteristic spectral information of soil organic carbon,reduce noise,and increase the correlation between SOC and reflectance.Compared with a single mathematical transformation,the simulation accuracy of multiple mathematical transformation combinations is further improved(such as CRFD-RF combination,R^(2)=0.87,RMSE=8.03,RPD=2.76;MSCFD-RF combination,R^(2)=0.87,RMSE=8.14,RPD=2.72).(3)Among the 4 simulation models,the overall simulation accuracy of random forest(RF)is the highest,among,CRFD-RF combination is the highest,and MSCFD-RF is the second.This study can provide a theoretical basis for SOC inversion and dynamic monitoring in Qinghai-Tibetan region using hyperspectral remote sensing.In the future,the formulation of different valid prediction models(covering several vegetation or soil types)is critical,as spectral characteristic band may vary from one type to another.
作者
周伟
曹鑫
王科明
肖洁芸
王婷
李浩然
姬翠翠
ZHOU Wei;CAO Xin;WANG Keming;XIAO Jieyun;WANG Ting;LI Haoran;JI Cuicui(Chongqing Jinfo Mountain Karst Ecosystem National Observation and Research Station,School of Geographical Sciences,Southwest University,Chongqing 400715,China;State Key Laboratory of Resources and Environmental Information System,Institute of Geographic Sciences and Natural Resources Research,Chinese Academy of Sciences,Beijing 100101,China;School of Smart City,Chongqing Jiaotong University,Chongqing 400074,China)
出处
《冰川冻土》
CSCD
北大核心
2023年第2期823-832,共10页
Journal of Glaciology and Geocryology
基金
国家自然科学基金项目(41501575,42171338)
中央高校基本科研业务费专项资金项目(SWU020015)
重庆市自然科学基金面上项目(cstc2021jcyj-msxmX0384)资助。
关键词
土壤有机碳
高光谱
机器学习
光谱变换
青藏高原
soil organic carbon
hyperspectral
machine learning
spectral transformation
Qinghai-Tibet Plateau