Quantitative analysis of heavy metals in soil via hierarchical deep neural networks with X-ray fluorescence spectroscopy

Abstract

Soil is an important source of potentially toxic metal intake for humans through the food chain. Accurate determination of elemental concentrations in soil is of great significance to safeguard human health, thereby requiring reliable detection techniques. This work aimed to propose a new quantitative analysis method that combines pre-processing and concentration prediction. A total of 59 soil samples were collected and subjected to laboratory analyses and energy dispersive X-ray fluorescence (ED-XRF) scans. Firstly, a novel pre-processing method based on iterative adaptive window empirical wavelet transform and Gaussian convolution (IAWEWT-GC) was proposed, which can identify and remove spectral background. Secondly, the instrument calibration curves of peak counts and elemental concentrations were established and the pre-treatment effect was verified by comparison with other common methods. The instrument calibration curves' coefficient of determination (R²) was enhanced to 0.988, 0.962, and 0.980 for Cr, Mn, and Cu, respectively. Finally, to accurately detect the elemental concentration, a hierarchical deep neural network (HDNN) was designed for estimating potentially toxic metals, enabling quantitative regression of multiple elements simultaneously. The proposed method was evaluated and compared with conventional algorithms (DNN, BPNN, and PLSR). The HDNN model achieved the highest R² of 0.943, 0.983, and 0.979 for Cr, Mn, and Cu, respectively, with the lowest RMSEP of 5.74, 84.45, and 29.17, and lowest MAE of 4.51, 78.51, and 16.35, indicating its ability to enable simple, rapid, and effective quantitative analysis of multiple elements. Overall, our proposed method provides a new option for the elemental analysis of samples rich in potentially toxic metals.