Estimating heavy metal concentrations in Technosols with reflectance spectroscopy

• Published in: Geoderma 2022-01, Vol.406, p.115512, Article 115512
• Main Authors: Kästner, Friederike, Sut-Lohmann, Magdalena, Ramezany, Shaghayegh, Raab, Thomas, Feilhauer, Hannes, Chabrillat, Sabine
• Format: Article
• Language: English
• Subjects: Partial Least Squares (PLS) regression; PTE; Random Forest regression; Reflectance Spectroscopy; Sewage farm; Soil environment monitoring
• ISSN: 0016-7061; 1872-6259
• DOI: 10.1016/j.geoderma.2021.115512

•VNIR-SWIR spectra were used to estimate PTEs (Cr, Cu, Ni and Zn) in Technosols.•Combining different soil preparation and spectral pre-processing steps.•Increasing soil preparation improves PLSR and RFR prediction accuracy.•Laboratory hyperspectral imagery provides good quality data for highly heterogeneous soil samples. Reflectance spectroscopy in the visible-infrared and shortwave infrared (450–2500 nm) wavelength region is a rapid, cost-effective and non-destructive method that can be used to monitor heavy metal (PTE, potential toxic elements) contaminated areas. Due to the PTE pollution that has accumulated in the course of wastewater treatment, the existence of Technosols presents an environmental problem, a potential source for PTE uptake by vegetation, or even the release of PTEs into groundwater. In this study, multivariate procedures using Partial Least Squares Regression (PLSR) and Random Forest Regression (RFR) are applied to quantify relationships between soil heavy metal concentration (Cr, Cu, Ni, Zn) and reflectance data of highly contaminated Technosols from a former sewage farm near Berlin, Germany. Laboratory measurements of 110 soil samples in four different preparation steps were acquired with HySpex hyperspectral cameras. The impact of the different preparation steps, namely “oven-dried”, “sieved”, “ground”, “LOI”, was evaluated for its potential to enhance the method performance or to reduce the time-consuming soil sample preparation. Furthermore, different spectral pre-processing methods were evaluated regarding improvements of spectral modelling performance and their ability to minimise noise and multiple scattering effects. Considering the optimal coefficient of determination (R2), PLSR shows an improving performance and accuracy with increasing preparation steps such as ground or LOI for all metals of interest (R2_Cr: 0.52–0.78; R2_Cu: 0.36–0.73; R2_Ni: 0.19–0.42 and R2_Zn: 0.41–0.74). RFR shows a weaker estimation performance for all metals, even when using higher sample preparation levels (R2_Cr: 0.36–0.62; R2_Cu: 0.17–0.72; R2_Ni: 0.20–0.35 and R2_Zn: 0.26–0.67). The results show that an application of methods such as PLSR for the prediction of PTE concentration in Technosols is still a challenge but provides more robust estimations than the user-friendly RFR method. Additionally, this study shows that PTE estimation performance in heterogeneous soil samples can be improved by increased laboratory soil preparation steps and furt