土壤有机质含量估测及其影响因素的光谱分析

 土壤颗粒大小差异使土壤反射光谱产生相应变化,影响土壤有机质含量等属性的光谱预测精度。本研究准备了颗粒粒径分别为2、0.25和0.15mm的土样,测定土壤有机质(Soil Organic Matter,SOM)含量,并于室内模拟条件下测定其反射光谱。通过分析不同粒径土样的原始(Raw)、多次散射校正(Multiple scattering correction, Msc)、一阶微分(First derivative, Fd)、连续统去除(Continuum removal, Cr)光谱与SOM含量之间的关系,筛选出与SOM含量相关性最强的Fd光谱单波段(2250nm, r=0.82, P<0.01),并建立线性回归模型;利用全波段光谱反射率,以偏最小二乘回归(Partial least square regression, PLSR)方法,确立2mm土样Msc处理光谱的PLSR模型为最优模型(RPD=3.56、R²=0.90、RMSEP=1.96g/kg)。土壤颗粒粒径对土壤光谱反射率变化有明显影响,但二者之间并非简单的线性关系,可能存在一个转折点;单变量(单波段光谱反射率)线性回归模型的预测能力,明显低于全波段反射光谱(Msc处理)-PLSR模型;土样样本容量对SOM含量预测精度有显著影响。因此,根据样本容量大小,选择合适的土壤颗粒粒径与光谱预处理方法组合可以提高预测精度。

Application of Spectroscopy in Estimating of Soil Organic Matter Content

Reflectance spectra of soil responds to the differences of soil particle sizes, which affects directly the capability of predicting soil organic matter （SOM） content by spectral reflectance. With soil samples of different particle sizes at 2mm, 0.25ram and 0. 15mm, reflectance spectra was collected under the con- dition of simulated sunshine in the laboratory and SOM contents were acquired. Multiple scattering correction （Msc）, first derivative （Fd） and continuum removal （Cr） were used as data pretreatment methods to raise the Signal-to-Noise of raw spectra （Raw） before analyzing the correlation between spectral reflec- tance and SOM contents. Single waveband at 2250nm is selected from Fd-treated reflectance spectra because of its maximal linear correlation coefficient （r=0.82, P〈0. 01） responding to SOM contents, and then a linear regression is built with a moderate precision （R2 =0. 69）. Furthermore, partial least square （PLS） was adopted to extract several principal components （PC） of full wavelength from 350 to 2500nm and to build prediction models. Only PLSR model of Msc-treated reflectance spectra of soil samples at 2mm particle size yielded better prediction for SOM content （RPD=3.56, R2 =0. 90, RMSEP= 1.96g/ kg）. It indicates that particle size of soil sample poses obvious effects on soil reflectance spectra. There is a conceivable turning point of particle size because the correlation between particle size and reflectance is not simple linear relation. Capability of predicting SOM content by univariate linear regression model is markedly lower than that of PLSR model. In addition, prediction precision is significantly affected by capacity of soil samples. Hence, satisfying prediction precision of SOM content can be acquired by effective combination of moderate particle size （2mm） and proper spectral pretreatment （Msc）.