传统线性回归模型在借助光谱信息进行土壤属性预测时,通常忽略了土壤自身所具有的空间异质性和依赖性,并且未考虑模型残差的空间结构。针对以上不足,该文以江汉平原232个土壤样本为研究对象,以土壤反射光谱为辅助变量,采用偏最小二乘回归、普通克里格、协同克里格以及回归克里格分别构建土壤有机碳密度预测模型,选取决定系数(R~2)、均方根误差、标准差与预测均方根误差比(ratio of performance to deviation,RPD)对模型预测精度进行对比评价。结果显示,结合高光谱信息,且同时考虑残差空间结构的回归克里格模型表现优于其他模型,预测决定系数R~2为0.617,RPD为1.614。鉴于土壤光谱信息同时还具有测定简单、省时、无损等优点,因此土壤光谱是土壤有机碳密度空间插值的理想辅助因子。
英文摘要:
The availability of soil organic carbon density (SOCD) information is of great importance for the development of ecological agriculture and the study of global climate change. Compared with traditional laboratory analysis, Visible and near-infrared (VNIR) reflectance spectroscopy has proven to be a rapid, non-destructive and cost effective method for estimating a variety of soil properties. It has got rapid development and has been applied in the field of soil research. In the prediction of soil properties by using spectral information, however, traditional linear regression models often ignore the spatial heterogeneity and dependency of soil, and fail to consider the spatial structure of the error term. With the aims to fill the current gap, a total of 232 topsoil samples were collected in Jianghan Plain with their spectral reflectance and SOCD measured. Partial least squares regression (PLSR), ordinary kriging (OK), co-kriging (COK), and regression kriging (RK) were used to estimate SOCD by using differently pre-treated spectral reflectance. Due to the facts that spectral pretreatments are crucial to reduce the physical variability and particle size effect, and are helpful to remove both additive and multiplicative effects in the spectra, five combinations of spectral pretreatments were utilized while predicting SOCD with PLSR. They were Savitzky-Golay smoothing (S-G), S-G + Multiplicative Scatter Correction (MSC), S-G+Standard Normal Variate (SNV), S-G + first derivative (1st), S-G + second derivative (2nd), S-G+1st+MSC, S-G+2~(nd)+MSC, S-G+1~(st)+SNV, S-G+2~(nd) +SNV. The prediction capabilities of the models were evaluated by R~2, root mean squared error (RMSE), and ratio of performance to deviation (RPD). Results showed that the RK approach which utilized soil spectra information outperformed the others, with the highest R~2-Pred 0.617 and RPD 1.614, and the lowest RMSEP 0.865 kg/m~2. PLSR took the second place with R~2-Pred 0.605, RPD 1.523 and RMSEP 0.917 kg/m~2, which was also acceptable for SOCD prediction. COK and OK generally failed in the predictions of SOCD, with R~2-Pred equaled to 0.007 and 0.004, RPD equaled to 0.903 and 0.874 and RMSEP equaled to 1.547 and 1.597 kg/m~2, respectively. Results indicated that the RK model, which considered both the spectral reflectance and the spatial structure of the error term of multivariate linear regression model can improve the prediction accuracy of SOCD. The fundamental reasons could be that soil spectra are comprehensive reflections of soil properties and those environmental factors that influence the formation of soil. Therefore, soil spectra are related with the variation of SOCD, and could be helpful in the prediction of SOCD. Besides, the optimal spectral pretreatment for PLSR modelling of SOCD is the combination of smoothing, first-order derivation and SNV. In summary, soil reflectance spectra in the visible and near-infrared region (350-2 500 nm) could serve as an effective proxy variable for SOCD estimation. Given that soil VNIR reflectance spectra are easy and quick to measure, and the measurement is also environmentally friendly, we would like to argue that soil spectral reflectance could serve as an ideal auxiliary variable for the spatial interpolation of SOCD.