Min Xiao,∗,∗∗ Qian Hu,∗,∗∗ Dong Ren,∗,∗∗ Anxiang Lu,∗,∗∗,∗∗∗ and Yi An∗∗∗∗
Heavy metal detection, spectral feature automatic selection, partialleast square
When using X-ray fluorescence (XRF) spectrometry to detect heavy metal content in soil, there are several factors that affect the range of spectral variable resolutions such as characteristic peak broadening of heavy metal, characteristic peak overlapping of heavy metal, instrumental noise, and potential absorption enhancement effect of elements. As a result, effective spectral variables of an element need to be detected and may be deviated from positions of its theoretical characteristic peak, further resulting in detection precision decline. In this study, the precision of the soil XRF spectral feature automatic selection method was improved, so that an optimal spectral region combination can be obtained through searching feature variables for the element, which could be detected in a full spectrum range based on FiPLS (Forward Interval Partial Least Square (PLS)) and BiPLS (Backward Interval PLS). An XRF spectrograph was utilized to scan spectra of 87 soil samples containing Cr in different concentrations. Then, the concentration of element Cr was measured for each sample by means of atomic absorption spectrometry (AAS). Finally, spectral characteristic variables were selected depending on FiPLS and BiPLS approaches. For the purpose of preventing the model from generating an overfitting effect, the optimal region count was located for spectra. Moreover, it was found that conditions were satisfactory when the total number of spectral regions was 30. Based on the above findings, a spectral quantitative model was established for Cr. RMSEP (Root Mean Square Error of Prediction) and correlation coefficients of results predicted by FiPLS, BiPLS, and PLS models were compared by means of design and experiment, which proved that FiPLS and BiPLS methods could improve model precision substantially. As indicated by relevant results, the feature variable automatic selection approach put forward here has the ability to get better performance of FiPLS and BiPLS in searching variable regions and extract spectral variables associated with elements to be detected in a more effective and more typical manner. Without a doubt, precision of the spectral quantitative model can be enhanced.
Important Links:
Go Back
