Estimation of multi-media metal(loid)s around abandoned mineral processing plants using hyperspectral technology and extreme learning machine.

Abstract

Hyperspectral techniques are promising alternatives to traditional methods of investigating potentially toxic metal(loid) contamination. In this study, hyperspectral technology combined with partial least squares regression (PLSR) and extreme learning machine (ELM) established estimation models to predict the contents of copper (Cu), zinc (Zn), arsenic (As), cadmium (Cd), lead (Pb) and tin (Sn) in multi-media environments (mine tailings, soils and sediments) surrounding abandoned mineral processing plants in a typical tin-polymetallic mineral agglomeration in Guangxi Autonomous Region. Four spectral preprocessing methods, Savitzky-Golay (SG) smoothing, continuum removal (CR), first derivative (FD) and continuous wavelet transform (CWT), were used to eliminate noise and highlight spectral features. The optimum combinations of spectral preprocessing and machine learning algorithms were explored, then the estimation models with best accuracy were obtained. CWT and CR were excellent spectral pretreatments for the hyperspectral data regardless of the applied algorithms. The coefficients of determination (R2) of estimation models for the best accuracy of various metals (loid) are as follows: Cu (CWT-ELM:0.85), Zn (CR-PLSR:0.93), As (CWT-ELM: 0.86), Cd (CR-PLSR: 0.89), Pb (CWT-PLSR: 0.75) and Sn (CR-ELM: 0.81). In contrast, ELM models had higher accuracy with R2 > 0.80 (except Cd and Pb). In conclusion, ELM-based spectral estimation models are able to predict metal (loid) concentrations with high accuracy and efficiency, providing a potential new combinatorial approach for estimating toxic metal contamination in multi-media environments.