基于惩罚最小二乘算法的土壤重金属检测光谱基线校正

针对X射线荧光光谱分析技术在检测土壤重金属过程中由于土壤背景复杂、包含大量噪声和干扰信息而易受基体效应影响的问题，为了提高定量分析模型的精度，利用惩罚最小二乘算法拟合基线与真实基线之间的保真度和平滑度，对X射线荧光光谱进行基线校正，从而减小基线漂移的影响。选用无基线扣除、非对称最小二乘(ASLS)、自适应迭代重加权惩罚最小二乘(AIRPLS)、非对称重加权惩罚最小二乘(ARPLS)、局部对称重加权惩罚最小二乘(LSRPLS)和多约束重加权惩罚最小二乘(DRPLS) 等6种处理方法对土壤重金属元素铅和砷的测量光谱进行基线校正;结合偏最小二乘(PLS)算法建立相应的校正模型，以选择最优基线校正算法;与神经网络(BP)和支持向量机(SVR)建立的校正模型进行比较，对模型进行评价。结果显示，铅元素的最佳模型为DRPLS-PLS，模型的R2达到0.982，预测均方根误差(RMSEP)为0.056 mg/kg;砷元素的最佳模型为DRPLS-PLS模型，模型的R2达到0.985，RMSEP为0.796mg/kg。与PLS和BP模型相比，铅、砷两种元素的SVR模型建模均最优，模型的R2分别达到0.998和0.993，RMSEP分别为0.015、0.596mg/kg。实验表明，通过基线校正后模型的预测精度、检出限和稳定性均有所提高，该方法可有效提高X射线荧光光谱在土壤中的定量分析能力。

Spectrum Baseline Correction for Soil Heavy Metal Detection Based on Penalized Least Squares Algorithm

X-ray fluorescence spectrometry has the advantages of nondestructive and rapid detection of heavy metals in soil. However, in the process of practical application, because the soil background is complex and contains a lot of noise and interference information, it is easy to be affected by the matrix effect. In order to improve the accuracy of the quantitative analysis model, it is necessary to carry out baseline correction of X-ray fluorescence spectrum and reduce the effect of baseline drift. Penalty least squares algorithm, as a common baseline algorithm, was used to further optimize the fitting baseline based on least squares by fitting the fidelity and smoothness between the baseline and the real baseline. No baseline deduction, asymmetric least squares (ASLS), adaptive iterative reweighted penalty least squares (AIRPLS), asymmetric reweighted penalty least squares (ARPLS), local symmetric reweighted penalty least squares (LSRPLS) and multi-constrained reweighted penalty least squares (DRPLS) were selected for baseline correction of the measured spectrum of heavy metal elements lead and arsenic in soil, and then the corresponding correction models were established with partial least squares (PLS) algorithm to select the optimal baseline correction algorithm. At last, the partial least square (PLS) model was compared with the correction model established by neural network (BP) and support vector machine (SVR) to evaluate the advantages and disadvantages of different models. The results showed that the optimal baseline correction algorithm of the two elements was DRPLS, which the R2 of the lead corresponding PLS model was 0.982, the prediction root mean square error (RMSEP) was 0.056 mg/kg, and the R2 of the arsenic corresponding PLS model was 0.985, the RMSEP was 0.796 mg/kg. Besides, the SVR models of lead and arsenic were optimal compared with PLS and BP models. And the R2 of the model reached 0.998 and 0.993, respectively. The RMSEP was 0.015 mg/kg and 0.596 mg/kg, respectively. Experiments showed that the prediction accuracy, detection limit and stability of the model established after baseline correction can effectively improve the quantitative analysis ability of X-ray fluorescence spectroscopy in soil.