消除水分因素影响的野外原状土壤盐分高光谱建模估测

土壤水分被确定为土壤属性(有机碳、盐分等)光谱预测准确性下降的一个主要原因,该文通过两种方法的对比旨在探索去除土壤水分影响、提高盐分高光谱定量估测精度的方法和技术路线。首先以山东省东营市垦利区为研究区,采用地物光谱仪测定了96个样本的野外原状土和室内风干土光谱,并进行一阶导数变换;接着,对比分析盐分光谱特征及水分的影响;然后分别采用外部参数正交化(external parameter orthogonalization,EPO)和非负矩阵分解(non-negative matrix factorizing,NMF)方法校正和融合野外原状土光谱,去除土壤水分因素的影响,形成野外原状土光谱的校正和融合光谱;最后基于野外原状土光谱、校正和融合光谱,分别采用多元逐步线性回归(multiple step linear regression,MSLR)和偏最小二乘回归(partial least squares regression,PLSR)构建土壤盐分含量的估测模型,并进行验证和比较,分析预测精度变化。结果显示:土壤水分对野外原状土光谱及盐分光谱特征影响较大,需要研究去除;EPO和NMF均可提高土壤盐分野外原位光谱估测精度,比较而言,NMF效果更为显著;EPO结合PLSR或NMF结合MSLR可作为去除水分影响的土壤盐分校准模型的技术路线。

Modeling and estimation of field undisturbed soil salt based on hyperspectra under removal of moisture factor

Abstract: Soil moisture is one of the main reasons for the decline of predictive accuracy of soil attributes (organic carbon, salt, etc.) in using spectrum analysis method. By comparing the two methods of external parameter orthogonization (EPO) and non-negative matrix factorization (NMF), the purpose of this article is to explore a method and technology route of removing the effect of soil moisture (SM) and improving the estimation precision of soil salt content (SSC) based on hyperspectra. Firstly, we used Kenli district of Dongying city in Shandong province as the research area, and took 96 soil samples in the fields. The samples'' hyperspectra in situ and indoor after air-dry were measured respectively by spectra radiometer, and then transformed to the first deviation. The content of soil salinity and moisture were measured in laboratory. Then, the spectral characteristics of soil salt content and the effect of soil moisture on it were analyzed by comparison. Next, the external parameter orthogonization and non-negative matrix factorization were respectively used to correct and fuse the soil spectra in situ (Situ-spectra), and to remove the SM effect, and form the EPO correction spectra (EPO-spectra) and the NMF fusion spectra (NMF-spectra) of the Situ-spectra. Finally the estimation models of SSC were built respectively by the multiple step linear regression (MSLR) and the partial least squares regression (PLSR) based on the Situ-spectra, EPO-spectra and NMF-spectra, and were verified and compared to analyze the change of the SSC prediction precision. The results indicated that the soil salt content was high the soil salt content gradient was obvious, and the dispersion degree of soil salt content was high. However the soil moisture content was about 30 times of the soil salt content in the study area. The correlation between soil salinity and spectra is better at the band ranges of 1440-1660 nm, 1830-1860 nm, 1960-2110 nm. The soil moisture had great effect on the Situ-spectra and soil salt content spectral characteristics. Therefore, it is necessary to remove soil moisture impact. The EPO method can reduce the correlation between spectra and soil moisture in most spectral regions, and at the same time weaken the correlation between spectra and soil salinity in local bands. In comparison, the NMF method can effectively reduce the correlation between spectra and soil moisture, and increase the correlation between spectra and soil salinity. Both the EPO and NMF can improve the accuracy of the soil salt content estimation based on Situ-spectra. After adoption of EPO, the validation coefficient of determination (R2) increased between 0.08 and 0.09, and the relative prediction deviation (RPD) increased between 0.08 and 0.69%. At the same time, after adoption of NMF, the validation R2 increased between 0.27 and 0.38 and reached above 0.80, the RPD increased between 1.04 and 1.06, and reached above of 2.37. Thus the result of NMF was more significant than that of EPO for the removal of the SM effect. The method of EPO combined with PLSR or NMF combined with MSLR can be used as the technical route of removing the soil moisture effect and building soil salt content calibration model. The results can effectively promote the quantitative remote sensing extraction and real-time, in-situ monitoring of the saline soil information.