基于室内高光谱数据的多种类型土壤有机质估算模型比较

基于高光谱数据的土壤有机质反演是土壤遥感及精准农业的重要研究内容,然而不同的光谱处理及建模方法使得模型的估算能力及精度差异明显,限制了模型之间的通用性。为了构建陕西省土壤有机质含量估算的最优模型,以陕西省9种主要土壤类型的216个土样的光谱反射曲线和土壤有机质含量为数据基础,将光谱反射曲线进行一阶微分d(R)、倒数对数log(1/R)、倒数对数一阶微分dlog(1/R)]和包络线去除N(R)4种变换,结合一元线性回归(SLR)、偏最小二乘回归(PLSR)和支持向量机回归(SVR)3种建模方法构建了不同的土壤有机质含量估算模型。结果显示:不同类型土壤的反射光谱曲线总体态势基本一致,吸收特征位置基本相同,且土壤有机质含量与光谱反射率呈负相关态势;基于d log(1/R)]光谱变换构建的SVR估算模型精度最高,建模集和验证集的判断系数(R~2)分别为0.9210、0.8874,验证均方根误差(RMSE)为2.18,相对分析误差(RPD)达到2.8751,是估算陕西省土壤有机质含量的最优模型,PLSR次之,SLR最差。

Comparison of multiple estimation models of soil organic matter based on laboratory hyperspectral reflectance in Shaanxi Province

Soil organic matter estimation based on hyperspectral data is an important research topic of soil remote sensing and precision agriculture. However, different spectral processing and modeling methods make the estimation ability and accuracy of models vary significantly, which limits the universality of models. In order to establish the optimal model of soil organic matter estimation in Shaanxi Province, based on the spectral reflectance curve and soil organic matter content of 216 soil samples of 9 main soil types in Shaanxi Province, the spectral reflectance curve was transformed into four transformations: first order differential d(R), reciprocal logarithm log(1/R), reciprocal logarithm first order differential d\log(1/R)\] and envelope line removal N(R). One variable linear regression (SLR), partial least square regression (PLSR) and support vector machine regression (SVR) were used to establish different models for estimating soil organic matter content. The results showed that the reflectance spectra curves and the absorption characteristic positions of different types of soil were basically the same, with the organic matter content and spectral reflectance showing a negative correlation trend. The accuracy of the nonlinear organic matter content estimation model based on SVR was the highest, PLSR was the second, SLR was the worst. Among them, the judgment coefficient R² of SVR model and validation model based on d\log(1/R)\] spectrum was the best, which were 0.9210 and 0.8874 respectively, the validation root mean square error RMSE was only 2.18, and RPD reached 2.8751. Therefore, SVR was the best model for estimating soil organic matter content.