Estimating potentially toxic elements in Brazilian tropical soils with vis-NIR spectroscopy and machine learning

• vis-NIR spectroscopy enables estimation of potentially toxic elements (PTEs). • Spectral pre-processing improved estimations, especially for Cr, Ni, Fe, and Ba. • Detrend and Savitzky-Golay pre-processing improve Cubist model estimations. • Continuum removal enhanced Fe and Pb estimations. • Creating spectral soil libraries for PTEs can be important for soil health. Traditional laboratory analyses for potentially toxic elements (PTEs) in soils are costly, time-consuming, and require chemical reagents. Visible and near-infrared (vis-NIR) reflectance spectroscopy offers a rapid, non-destructive, and cost-effective complementary tool that can reduce the number of traditional laboratory analyses and consequently the use of chemical reagents, mainly when supported by robust spectral libraries. This study aimed to (i) assess the potential of vis-NIR spectroscopy to quantify natural concentrations of Ba, Cr, Cu, Fe, Mo, Ni, Pb, and V in tropical soils, (ii) evaluate spectral pre-processing procedures combined with the Cubist model, and (iii) identify the most relevant spectral bands for estimating these elements. A total of 242 soil samples (0-20 cm depth) were analysed for Ba, Cr, Cu, Fe, Mo, Ni, Pb and V using ICP-OES. Spectral reflectance (350-2500 nm) was acquired and pre-processed with Savitzky-Golay (SG), standard normal variate (SNV), detrend (DET) and continuum removal (CR). Estimations were modelled using the Cubist algorithm. Pre-processing spectra improved estimations compared to raw spectra. The Cubist model combined with Savitzky-Golay enhanced estimations for Ba, Cu, and Mo. Detrend was most effective for Cr, Ni, and V and continuum removal improved Fe and Pb. Reasonable performance was achieved for Ba (R 2 adj =0.55; RPIQ=1.40), Cr (R 2 adj =0.65; RPIQ=1.91), and Ni (R 2 adj =0.70; RPIQ=1.60). Vis-NIR spectroscopy coupled with machine learning shows potential for estimating certain PTEs in highly variable tropical soils. While predictions for Ba, Cr, and Ni were promising, other elements require larger datasets or alternative modelling approaches.