土壤光谱特征是土壤内在理化特性的外在表现, 利用可见-近红外(Vis-NIR)的漫反射光谱估测土壤有机质含量(SOMC)可为土地资源的合理开发与利用提供重要的科学依据。 土壤是由多种物质组成的混合物, 其高光谱数据中存在某些成分(例如盐颗粒和矿物质)的重叠吸收, 同时波段之间存在共线性问题, 这些因素对光谱分析和建模带来了极大的挑战。 光谱指数法通过迭代运算, 不但充分考虑波段之间的协同作用, 而且具有最小化无关波长影响的功能。 此外该方法将光谱特征从一维扩展到多维, 能容易地检测和区分出细微的吸收峰。 以新疆艾比湖流域为研究区, 采集了120个土壤样品, 在室内进行SOMC和光谱的测定。 利用一阶微分(FD)和连续统去除(CR)对高光谱数据进行预处理。 在已有两波段指数的基础上, 加入第三个波段, 利用最优波段算法, 构建了三种SOMC的三波段光谱指数(TBI), 并从光谱机理上讨论了TBI的合理性。 最后根据支持向量机(SVM)的建模效果, 进一步比较不同维度光谱参数对SOMC估测的准确性。 研究结果表明: (1)光谱预处理技术可以在一定程度上减弱反射光谱中的噪声信息, 突出更多潜在的光谱信息; (2)通过对比分析得出, SOMC的相关性随着光谱信息维度的增加而增加, 即, TBI>二波段指数>一维光谱参数; (3)新开发的TBI在SOMC的建模和验证过程中提供了比两波段指数和一维光谱参数更好的估测效果, 其中TBI-1的估测效果最好, 建模集的决定系数(R2C)为0.88, 验证集的决定系数(R2V)为0.85, 相对分析误差(RPD)为2.43。 该研究对比了不同维度光谱参数对SOMC的响应和建模精度, 发现三波段光谱指数是评价SOMC的重要参量。 此外, TBI与SVM算法的结合, 可以显著弱化土壤噪声信息, 提高SOMC的预测精度, 在土壤其他生化参数的估计中具有较强的应用潜力。

The spectral characteristics of soil are the external manifestation of physical and chemical properties in soil. Estimating soil organic matter content (SOMC) by visible-near infrared (VIS-NIR) diffuse reflectance spectroscopy could provide an important scientific basis for the rational development and utilization of land resources. However, the soil is a mixture of many substances, and its hyperspectral data have overlapping absorption of certain components (such as salt particles and minerals), and there are collinear problems between the bands, which bring great challenges for spectral analysis and modeling. Through the iterative operation, the spectral index method not only fully consider the synergy between the bands, but also has the function of minimizing the influence of irrelevant wavelengths. In addition, the method extends the spectral features from one dimension to multidimensional, and can easily detect and distinguish subtle absorption peak. In this study, 120 soil samples were collected from the Ebinur Lake Basin in Xinjiang, and SOMC and spectra were measured indoors. Hyperspectral data were preprocessed using first derivative (FD) and continuum removal (CR). Based on the existing two-band index, the third band was added, and the three-band spectral index (TBI) of three SOMCs was constructed by using the optimal band algorithm. The rationality of TBI was discussed from the spectral mechanism. Finally, according to the modeling effect of support vector machine (SVM), the accuracy of SOMC estimation by different dimensional spectral parameters was further compared. The research results showed that: (1) Spectral pretreatment technology could weaken the noise information in the reflection spectrum to some extent and highlighted more potential spectral information; (2) Through comparative analysis, the correlation of SOMC increased with the increase of the spectral information dimension, that was, TBI>two-band index>one-dimensional spectral parameters; (3) The newly developed TBI provided better estimation results than the two-band index and one-dimensional spectral parameters in the modeling and verification process of SOMC. The TBI-1 had the best estimation effect and the determination coefficient of the modeling set. (R2C) was 0.88, the decision coefficient (R2V) of the verification set was 0.85, and the relative analysis error (RPD) was 2.43. In summary, this study compared the response and modeling accuracy of different dimensional spectral parameters to SOMC. It was found that the three-band spectral index was an important parameter for evaluating SOMC and had good performance. In addition, the combination of TBI and SVM algorithm could weaken soil noise information, improved the prediction accuracy of SOMC, and had strong application potential in the estimation of other biochemical parameters of soil.