土壤有机碳（SOC）对土壤肥力至关重要, 可见-近红外光谱能对其实现快速反演, 为区域监测和定量遥感提供基础。 针对包络去除（CR）仅提供反射光谱的单向吸收特征, 多元回归中预测信息缺失、 拟合结果未充分反映波段特征, 利用世界土壤数据库245份中国土样的可见-近红外光谱, 首次提出双包络去除（BCR）与正交偏最小二乘（OPLS）结合的反演方法BCR-OPLS, 同时纳入光谱反射率及上、 下边包络去除量, 讨论组分参考值偏态分布时幂函数或对数缩放在回归时的优化作用, 建立多种土壤的综合与分类估计模型, 并导出适用特定类型土壤的SOC指数。 结果表明, 对多种土壤有机碳含量反演, 相较PLSR模型（决定系数R2和估计根均方误差RMSEE分别为0.69和0.45%）, BCR-OPLS模型的预测能力明显改善（R2和RMSEE分别为0.9和0.26%）; 而对单一类型土壤的反演精度则进一步提升, 根据载荷趋势和变量重要性建立的SOC指数, 预测如黄色铁铝土的有机碳含量时（以400, 590和920 nm）, 其反演结果R2达到0.94、 RMSEE达到0.21%。 双包络去除与OPLS相结合, 增强了光谱特征诊断的鲁棒性, 提高了不同类型土壤的综合与分类SOC全谱反演精度, 基于直观的图谱表达可构建简单的波段预测关系, 深化了物理经验吸收与统计多元回归之间的联系。

Soil Organic Carbon (SOC) is important for soil fertility and can be quickly retrieved by Visible Near-Infrared (VNIR) Spectroscopy, which provides a basis for regional monitoring and quantitative remote sensing. For the traditional Continuum Removal (CR) method, only the upside absorption characteristics of the reflection spectrum envelope is considered in multiple regression, which results in the absence of CR downside or predictive spectral background information, thus the variables usually do not reflect the emission characteristics of all band . In this paper, a new method named BCR-OPLS which combines Bi-Continuum Removal (BCR) and Orthogonal Partial Least-Squares (OPLS) is proposed for SOC content retrieval, conducting a test upon 245 Chinese soil samples containing VNIR (350~2 500 nm) diffuse reflectance spectra downloaded from ICRAF-ISRI Database. With BCR-OPLS method, both the upside and downside continuum removal are included in analyzing the characteristics of the spectra. After building the comprehensive and classification model for soils of different types mixed and alone, an SOC index applicable to certain type of soil is derived. The role of power function and logarithmic function playing in skewness correction for the SOC reference values' statistical distribution is discussed. As a result, by introducing bilateral-continuum information, the SOC retrieval ability of the BCR-OPLS model is significantly improved (Coefficients of determination R2=0.9 and Root mean square error Estimated RMSEE=0.26%) compared with the initial R-PLSR model (R2=0.69, RMSEE=0.45%), and the SOC retrieval accuracy of a certain type is further improved. For example, when predicting SOC of the Orthic Ferralsols (using 400, 590 and 920 nm), R2 and RMSEE improved to be 0.94 and 0.21% respectively. In summary, BCR-OPLS enhances the robustness of spectral feature diagnostics by improving the accuracy of both comprehensive and classified SOC inversion based on full-spectrum, and derives a simple SOC prediction index composed of several wavelength variables for a certain type of soil through the translatability of relationships among BCR and SOC content revealed in loading scatter plot of OPLS, which are selected according to the loadings' trend and Variable Importance in Projection. Finally, BCR-OPLS strengthens the connection between experienced physical absorption analysis and obscure statistical multiple regression method.