基于NMF-PLS对含水量影响下土壤重金属含量反演模型研究

土壤中过高的重金属含量危害巨大，不仅造成了严重的环境污染，而且通过食物链进入人体对人体健康造成严重威胁，所以对重金属检测十分重要。X射线荧光光谱法具有检测时间短、无损检测、检测成本低等特点被广泛使用，然而检测的光谱数据因受到土壤含水量因素的严重干扰，导致直接对土壤重金属含量估算精度较低。以河北省保定市满城区土样为研究对象，对采集的土样进行除杂、过筛、烘干后加入一定量重金属溶液制备不同含水量不同重金属的样本进行检测。对实验中异常数据计算了马氏距离和进行NJW聚类予以剔除，分析了土壤含水量对土壤重金属光谱的影响，结果表明不同含水量间光谱重复性差，随着土壤含水量的增加光谱强度呈非线性降低。采用Savitzky-Golay卷积平滑去噪法和线性本底法对光谱进行预处理，以解决因环境、仪器本身带来的噪声和基线漂移等问题。然后针对于土壤含水量这一主要干扰，采用非负矩阵分解算法进行处理，并使用峰值信噪比这一评价模型确定端元数目，结果表明当端元数目增至10时峰值信噪比趋于稳定波动很小，非负矩阵分解处理后相同重金属含量不同含水量间光谱重复性好、相似性好，并计算了光谱间的相关系数进一步证明了光谱间的相似性。去除含水量对于光谱干扰后建立了偏最小二乘法预测模型，为了验证预测模型的精度，建立了未去除含水量的偏最小二乘法预测模型和使用外部参数正交化法去除含水量建立的偏最小二乘法预测模型，并使用评价参数决定系数(R2)、交叉验证均方根误差(RMSECV)、平均绝对误差(MAE)和相对分析误差(RPD)进行评价。验证结果表明，相比较未去除含水量建立的模型，使用非负矩阵分解去除含水量建立的偏最小二乘法模型R2和RPD分别提高了0.019 7和1.029 2，RMSECV和MAE分别降低了2.386 3和1.439 6；相对于外部参数正交化法建立的偏最小二乘法模型，R2和RPD分别提高了0.009 9和0.108 1，RMSECV和MAE分别降低了0.244 7和0.356 6，说明了经过非负矩阵分解去噪后建立的模型有效提高了预测的精度和鲁棒性。非负矩阵分解可以有效消除土壤含水量对光谱的影响，在此基础上建立的偏最小二乘法模型实现了土壤重金属含量的反演，为重金属定量检测提供了一定的技术支持。

Study on Inversion Model of Soil Heavy Metal Content Based on NMF-PLS Water Content

The excessively high content of heavy metals in the soil is hugely harmful,not only causing serious environmental pollution,but entering the human body through the food chain poses a serious threat to human health,so it is very important for heavy metal detection.X-ray fluorescence spectroscopy has been widely used because of its short detection time,non-destructive testing,and low testing costs.However,the detection of spectral data is severely disturbed by soil moisture factors,which leads to lower accuracy in estimating the heavy metal content in the soil directly.Taking the soil samples of Mancheng District,Baoding City,Hebei Province as the research object,the collected soil samples were cleaned,screened,dried,and then added with a certain amount of heavy metal solution to prepare samples with different water content and heavy metals for detection.The Mahalanobis distance and NJW clustering were calculated for the abnormal data in the experiment,and the influence of soil moisture content on the heavy metal spectrum was analyzed,the results show that the spectral repeatability of different water content is poor,and the spectral intensity decreases nonlinearly with the increase of soil water content.The Savitzky-Golay convolution smoothing denoising method and linear background method are used to preprocess the spectrum to solve the problems of noise and baseline drift caused by the environment and the instrument itself.A non-negative matrix factorization algorithm was used to deal with the peak signal-to-noise ratio evaluation model to determine the number of end elements.The results show that the peak signal-to-noise ratio tends to increase when the number of end elements increases to 10.The stable fluctuation is very small.After the non-negative matrix decomposition treatment,the spectrum repeatability and similarity are good among the same heavy metal content and different water content.The correlation coefficient between the spectra is calculated to prove the similarity between the spectra further.A partial least squares prediction model was established after removing the water content for spectral interference.In order to verify the accuracy of the prediction model,a partial least squares prediction model with no water content removed was established,and the partial water content was removed by orthogonalization with external parameters The least squares prediction model is evaluated using the evaluation parameter determination coefficient(R^2),cross-validated root mean square error(RMSECV),average absolute error(MAE),and relative analysis error(RPD).Validation results show that compared to models built without removing water content,non-negative moments are used partial least squares model established by matrix decomposition and removal of water content R^2 and RPD increased by 0.0197 and 1.0292,RMSECV and MAE decreased by 2.3863 and 1.4396;Compared to the partial least squares model established by the external parameter orthogonalization method,the RPD and RPD increased by 0.0099 and 0.1081,and the RMSECV and MAE decreased by 0.2447 and 0.3566,it is shown that the model established after denoising by non-negative matrix decomposition can effectively improve the accuracy and robustness of prediction.Non-negative matrix factorization can effectively eliminate the effect of soil water content on the spectrum,and the partial least squares model established on this basis has realized the inversion of soil heavy gold content and provided certain technical support for quantitative detection of heavy metals.