针对土壤中铅含量的定量检测问题, 本研究基于太赫兹光谱技术对不同pH下土壤中铅含量的最佳反演预测模型进行了探索性研究。 分别制备了pH为8.5, 7.0和5.5的含铅土壤样品, 采集样品的太赫兹光谱数据, 并对光谱数据做了多元散射矫正(MSC)、 基线校正和Savitzky-Golay平滑等预处理。 对预处理后的光谱数据, 采用连续投影法(SPA)选取光谱数据的特征频率。 基于选取的特征频率分别采用偏最小二乘法(PLS)、 支持向量机(SVM)和误差反向传播神经网络(BPNN)建立土壤中铅含量的反演预测模型, 采用校正集相关系数(Rc)、 校正集均方根误差(RMSEC)、 预测集相关系数(Rp)、 预测集均方根误差(RMSEP)和剩余预测偏差(RPD)作为评价参数对模型性能进行评估, 确定铅在不同pH土壤中的最佳预测模型。 实验结果表明: 在经过SPA选择特征频率后的建模效果普遍比全光谱的效果好。 其中pH 8.5的样品最佳预测模型为SPA-PLS, Rc, Rp, RMSEC, RMSEP和RPD分别为0.997 7, 0.994 6, 14.52 mg·kg-1, 22.70 mg·kg-1和9.63; pH 7.0的样品最佳预测模型为SPA-SVM, Rc, Rp, RMSEC, RMSEP和RPD分别为0.996 2, 0.975 7, 20.25 mg·kg-1, 33.04 mg·kg-1和4.56; pH 5.5的样品最佳预测模型为SPA-BPNN, Rc, Rp, RMSEC, RMSEP和RPD分别为0.968 7, 0.974 4, 48.83 mg·kg-1, 55.03 mg·kg-1和4.44。 该研究结果为不同pH土壤中铅含量的光谱反演预测提供了一种新思路, 亦可为其他重金属在不同pH土壤中的含量反演预测模型提供理论方法和技术支持。

Aiming at the quantitative determination of heavy metal lead in soils, the optimal inversion prediction model of lead content in soils at different pH was studied based on terahertz spectroscopy. Lead-containing soil samples with pH of 8.5, 7.0 and 5.5 were prepared. Terahertz time-domain spectroscopy system TERA K15 was used to collect the Terahertz spectra of the samples, and multivariate scattering correction (MSC), baseline correction and Savoitzky-Golay smoothing were used to pre-process the spectra. For the spectral data of pre-treatment, successive projection algorithm (SPA) was used to select the sensitive frequencies of spectral data. Based on the selected sensitive frequencies, partial least squares (PLS), support vector machine (SVM) and back propagation neural network (BPNN) was used to establish inversion prediction models of lead content in the soil. The correlation coefficient of calibration (Rc), root mean square error of calibration (RMSEC), the correlation coefficient of prediction (Rp), root mean square error of prediction (RMSEP) and residual predictive deviation (RPD) were used as model evaluation parameters to evaluate the performance of the model, and to determine the best prediction model of leadship in different pH soils. The experimental results show that the modeling effect after SPA choosing sensitive frequencies is generally better than that of full spectrum. Among them, the best prediction models for the samples with pH 8. 5 were SPA-PLS, Rc, Rp, RMSEC, RMSEP and RPD were 0.997 7, 0.994 6, 14.52 mg·kg-1, 22.70 mg·kg-1 and 9.63, respectively; the best prediction models for the samples with pH 7.0 were SPA-SVM, Rc, Rp, RMSEC, RMSEP and RPD were 0.996 2, 0.975 7, 20.25 mg·kg-1, 33.04 mg·kg-1 and 4.56, respectively; and the samples with pH 5.5 were the best. The prediction models are SPA-BPNN, Rc, Rp, RMSEC, RMSEP and RPD are 0.968 7, 0.974 4, 48.83 mg·kg-1, 55.03 mg·kg-1 and 4.44, respectively. The results provide a new idea for inversion prediction of lead content in different pH soils, and also provide theoretical methods and technical support for other heavy metals inversion prediction models in different pH soils.