基于土壤协变量与VIS-NIR光谱估算土壤有机质含量的研究

Soil Organic Matter Content Estimation Based on Soil Covariate and VIS-NIR Spectroscopy

为研究土壤有机质(SOM)含量与土壤电导率(EC),pH和Fe的相关关系,立足于艾比湖保护区,在2017年8月共收集了110个样本,测量了土壤反射光谱、SOM含量、土壤协变量(EC,Fe,pH)。对原始光谱进行了三种预处理:SG(Savitzky-Golay)平滑、多元散射校正(MSC)和一阶微分(FD),并对光谱数据进行了主成分分析(PCA),选取前5个主成分(PC)的特征值作为光谱变量。以使用原始光谱数据、两种预处理方法(SG-MSC、SG-MSC-FD)作为策略Ⅰ,以土壤协变量(EC,Fe,pH)为预测变量作为策略Ⅱ,以策略Ⅰ和策略Ⅱ组合作为策略Ⅲ,分别利用偏最小二乘回归(PLSR)建立SOM的预测模型。结果表明,基于预处理后的光谱数据的预测效果(验证集中决定系数为R2=0.66~0.82)优于以土壤协变量为预测变量的预测效果(验证集中R2=0.40),此外将土壤协变量与光谱数据相结合可以明显改善SOM的光谱预测精度(最佳验证集中R2=0.88)。同时,对光谱数据进行预处理后,能够有效增强潜在的光谱信息,提高模型的预测精度。综上,将可见光-近红外光谱信息和土壤协变量相结合的策略能够有效提升SOM模型的预测性能。

To investigate the relationship of the soil organic matter(SOM)content to the electrical conductivity(EC),pH,and Fe content,we collected 110 samples at the Ebinur Lake Reserve in August 2017 and measured the soil reflectance spectra,SOM content,EC,Fe content,and pH.We performed three kinds of pre-treatments,including Savitzky-Golay(SG)smoothing,multiplicative scatter correction(MSC),and first-order differentiation(FD),on the original spectrum and then performed a principal-component analysis of the spectral data.The eigenvalues of the first five principal components were selected as the spectral variables.Strategy Ⅰ used the original spectrum,performed SG-MSC and SG-MSC-FD on it,and employed the original spectrum as a control group.Strategy Ⅱ used the soil covariates(EC,Fe,pH)as the input variables.Strategy Ⅲ combined strategy Ⅰ and strategy Ⅱ.Predictions of the SOM content were obtained for all three strategies using partial least squares regression.The results show that predictions based on the pre-processed spectral data(for the verification set,the coefficient of determination was R^2=0.66-0.82)were better than those based on the soil covariates as the prediction variables(for the verification set,the coefficient of determination was R^2=0.40)and that combining the soil covariates and spectral data significantly improved the spectral-prediction accuracy for SOM(for the best verification set,R^2=0.88).Pre-processing the spectral data effectively enhanced the potential spectral information and improved the predictive accuracy of the model.In summary,the combination of visible-near-infrared spectral information and soil covariates effectively improves the predictive performance of SOM models.