摘要
【目的】快速、准确地监测土壤有机质对于精准农业的发展具有重要意义。可见光-近红外(visible and near-infrared,Vis-NIR)光谱技术在土壤属性估算、数字化土壤制图等方面应用较为广泛,然而,在田间进行光谱测量,易受土壤含水量(soil moisture,SM)、温度、土壤表面状况等因素的影响,导致光谱信息中包含大量干扰信息,其中,SM变化是影响光谱观测结果最为显著的因素之一。此研究的目的是探讨OSC算法消除其影响,提升Vis-NIR光谱定量估算土壤有机质(soil organic matter,SOM)的精度。【方法】以江汉平原公安县和潜江市为研究区域,采集217份耕层(0—20 cm)土壤样本,进行风干、研磨、过筛等处理,采用重铬酸钾-外加热法测定SOM;将总体样本划分为3个互不重叠的样本集:建模集S^0(122个样本)、训练集S^1(60个样本)、验证集S^2(35个样本);设计SM梯度试验(梯度间隔为4%),在实验室内获取S^1和S^2样本集的9个梯度SM(0%—32%)的土壤光谱数据;分析SM对土壤Vis-NIR光谱反射率的影响,采用外部参数正交化算法(external parameter orthogonalization,EPO)、正交信号校正算法(orthogonal signal correction,OSC)消除SM对土壤光谱的干扰;利用主成分分析(principal component analysis,PCA)的前两个主成分得分和光谱相关系数两种方法检验消除SM干扰前、后的效果;基于偏最小二乘回归(partial least squares regression,PLSR)方法建立EPO和OSC处理前、后的SOM估算模型,利用决定系数(coefficient of determination,R^2)、均方根误差(root mean square error,RMSE)和RPD(the ratio of prediction to deviation)3个指标比较PLSR、EPO-PLSR、OSC-PLSR模型的性能。【结果】土壤Vis-NIR光谱受SM的影响十分明显,随着SM的增加,土壤光谱反射率呈非线性降低趋势。OSC处理前的湿土光谱数据主成分得分散点相对分散,与干土光谱数据主成分得分空间的位置不重叠,不同SM梯度之间的光谱相关系数变化较大;OSC处理后的湿土光谱数据主成分得分空间的位置基本与干土光谱数据相重合,各样本光谱数据之间相似性很高,不同SM梯度之间的光谱相关系数变化较小。9个SM梯度的EPO-PLSR模型的验证平均R^2_(pre)、RPD分别为0.69、1.7。9个SM梯度的OSC-PLSR模型的验证平均R^2_(pre)、RPD分别为0.72、1.89,校正后的OSC-PLSR模型受SM的较小,有效提升SOM估算模型的精度和鲁棒性。【结论】OSC能够消除SM变化对土壤Vis-NIR光谱的影响,可为将来田间原位实时监测SOM信息提供一定的理论支撑。
【Objective】 Rapid and accurate quantitative analysis of soil organic matter is essential for sustainable development of precision agriculture. Visible and near-infrared（Vis-NIR） reflectance spectroscopy has been widely used for soil properties estimation and digital soil mapping. However, it is less exact in monitoring soil organic matter（SOM） in the field when compared to laboratory-based spectroscopic measurement mainly due to some factors, such as soil moisture, temperature, and soil surface texture. Among these three factors, soil moisture（SM） has the most pronounced effects on spectral reflectance. Therefore, it is urgently significant that a method for removing SM effects from spectral reflectance and improving the accuracy of quantitative prediction of SOM should be proposed. 【Method】 A total of 217 soil samples used in this study were collected at 0-20 cm depth from Gong＇an County and Qianjiang City in Jianghan Plain. These soil samples were air-dried, ground, and sieved（less than 2 mm） in the laboratory, and the SOM of each soil sample was analyzed based on potassium dichromate external heating method. These 217 samples were further divided into three non-overlapping data-sets： The model calibration set（S^0）, this set consisted of 122 samples to develop multivariate models for SOM; The orthogonal signal correction（OSC） development set（S^1）, this set consisted of 60 samples for OSC development; The validation set（S^2）, this set consisted of 35 samples for independent OSC validation. Then, sample rewetting（S^1 and S^2 set） was carried out： each soil sample was weighed 150 g oven-dried soil in a cylindrical black box, and then they were rewetted by 4% SM increment for each level in the laboratory. Total 9 treatments were obtained, corresponding to the following SM levels i.e. 0%, 4%, 8%, 12%, 16%, 20%, 24%, 28%, and 32%. Soil hyperspectral reflectance was measured in the laboratory with an ASD Fieldspec-Pro spectroradiometer for the three data-sets（S^0, S^1 and S^2, including the rewetting samples）. Savitzky-Golay smoothing with a window size of 11 nm and polynomial order of 2（SG） were applied to the three data-sets, then external parameter orthogonalization（EPO） and orthogonal signal correction（OSC） were conducted to remove the SM effects on reflectance spectra. In the next, the effect of SM on the reflectance spectra was analyzed, and the scores of the first two principal components from the principal component analysis（PCA） corrected by OSC method and spectral correlation coefficient were used to compare the performance in removing the effects of SM. Finally, the S^0 data-sets were calibrated using the partial least squares regression（PLSR）, and the S^2 data-sets were then examined as external validation sets. Using the coefficient of determination（R^2）, root mean squared error（RMSE） and the ratio of prediction to deviation（RPD） between the predicted and measured SOM to compare the performance of PLSR, EPO-PLSR and OSC-PLSR models, high R^2, RPD and low RMSE were indicators of the optimal model in the removal of SM effects.【Result】SM had an obvious influence on soil spectra reflectance, and the reflectance values across the entire wavelength domain decreased as the SM increased, making it more challenging to identify useful features of SOM by spectra, it dramatically degraded the prediction accuracy of SOM. No overlap before OSC was observed between the wet and dry ground spectra because the wet spectra grouped in an independent space from the dry ground spectra, and the range of the spectral correlation coefficients between different SM levels was large. However, after OSC, the wet spectra had nearly identical positions in the feature space to the corresponding dry ground spectra, which showed the spectral similarity between the two groups of spectra, and the range of the spectral correlation coefficients between different SM levels was small. The validation mean values of R^2 pre, RPD for the nine SM levels of EPO-PLSR model were 0.72 and 1.89, respectively. OSC method could effectively remove the effects of SM on SOM estimation, OSC-PLSR model obtained a better performance than the PLSR, EPO-PLSR model, the validation mean values of R2 pre, RPD for the nine SM levels were 0.72 and 1.89, respectively. 【Conclusion】 OSC-PLSR method was recommended for a better quantitative prediction of SOM from the soil samples under different SM levels. In the future, this approach may facilitate the proximally sensed field spectra for rapidly measuring SOM for this study area.
出处
《中国农业科学》
CAS
CSCD
北大核心
2017年第19期3766-3777,共12页
Scientia Agricultura Sinica
基金
国家自然科学基金(41401232)
中央高校基本科研业务费专项资金(CCNU15A05006)
华中师范大学研究生教育创新资助项目(2016CXZZ15)
关键词
Vis-NIR光谱
土壤有机质
土壤含水量
正交信号校正
偏最小二乘回归
江汉平原
Vis-NIR spectra
soil organic matter （SOM）
soil moisture （SM）
orthogonal signal correction （OSC）
partial leastsquares regression （PLSR）
Jianghan Plain
