A methodology to estimate REE contents in post-mining soils by diffuse reflectance spectroscopy and machine learning

Abstract

Here we evaluated the capacity of diffuse reflectance spectroscopy (DRS) to predict the content of rare earth elements (REEs), namely Y, La, Ce, Pr, and Nd, in post-mining soils susceptible to be revisited for secondary mining. This technology was combined with machine learning (ML) data processing to achieve REE measurements using a non-destructive, rapid, and portable methodology as an alternative to geochemical surveys. A former mining site with a notable variability of substrates was selected to demonstrate the methodology. Spectral reflectance was recorded at 1-nm intervals from Visible and Near-Infrared (VIS-NIR), 350–2500 nm, and classical geochemical data were achieved to evaluate the performance of DRS. The REE content was adjusted by optimizing sample preparation (sieving at 2 mm was found to be the most adequate approach), transformation of spectral data (first-derivative transformation was selected), and the most appropriate ML technique to obtain fitted models (non-parametric techniques yielded models with a better fit than parametric ones). Random Forest (RF) provided non-biased models for La, Ce, Pr, and Nd and Generalized Boosting Models (GBMs) for Y, with R² and rRMSE ranging from 0.840 to 0.890 and from 10.98 to 6.74, respectively. The models with the best performance revealed that NIR spectral region, specifically long-wave near infrared (LNIR), was the most important for determining REE content. Also, some wavelength ranges within the spectrum were identified for each REE. This reliable methodology emerges as a rapid approach to evaluate REE concentrations in secondary mining soils and has possible applications in primary mining exploration, environmental remediation, metallurgy, and materials science.