Study on LOD of Trace Elements by XRF Analysis Using BP & Adaboost and PLS Methods

Abstract

Poisonous elements such as Cd, Hg, Pb, As, Zn, Cr, Ni, Cu etc. are commonly observed in polluted soil and hard to be removed by soil microbes. It is of significant importance to identify these poisonous elements in-situ and accurately both in qualitative and quantitative sense. In order to determine the Limit of Detection (LOD) for trace elements (e.g. Cadmium) in polluted soil samples based on Energy Dispersion X-ray fluorescence (ED-XRF) spectroscopy, approximately 60 national standard soil samples were collected and measured by an XRF equipment. The authors firstly utilize the Method Detection Limit (MDL) algorithm to calculate the LOD of trace elements, and then develop a new model called Back Propagation Adaboost (BP & Adaboost) classification to determine the LOD based on a presumed tolerance error (e.g. 5%). Furthermore, the Multivariate- Partial Linear Squares Regression (M-PLSR) method is applied to regress the data and validate the LOD values. In this paper, the authors make a detailed comparison between the BP algorithm and the BP & Adaboost classification algorithm under different presumed detection limits, and it is found that the detection results achieved the best qualitative prediction of Cd element (i.e. whether it exists in soil) based on the BP & Adaboost algorithm. The experimental results indicate that the BP & Adaboost algorithm is the most effective method to determine and decrease the LOD of trace element (such as Cd) in soil. The advantages are: It combines the classification effects of several weak classifiers, and determines that the LOD of element Cd is 0.5mg/kg with prediction error rate of 5%. Compared with the traditional methods like MDL, it is proved that the BP & Adaboost algorithm is appropriate to be used in the terms of prediction accuracy. It is recommended that the BP & Adaboost classification method shall be used for material analysis on XRF spectroscopy.