Remediation-related monitoring of heavy metal concentration of contaminated Technosol using hyperspectral measurements

Across Europe there are 2.5 million potentially contaminated sites, approximately one third have already been identified and around 15% have been sanitized. Phytoremediation is a well-established technique to tackle this problem and to rehabilitate soil. However, remediation methods, such as biological treatments with microorganisms or phytoremediation with trees, are still relatively time consuming. A fast monitoring of changes in heavy metal content over time in contaminated soils with hyperspectral spectroscopy is one of the first key factors to improve and control existing bioremediation methods.

At former sewage farms near Ragow (Brandenburg, Germany), 110 soil samples with different contamination levels were taken at a depth between 15-20 cm. These samples were prepared for hyperspectral measurements using the HySpex system under laboratory conditions, combing a VNIR (400-1000 nm) and a SWIR (1000-2500 nm) line-scan detector. Different spectral pre-processing methods, including continuum removal, first and second derivatives, standard normal variate, normalisation and multiplicative scatter correction, with two established estimation models such as Partial Least Squares Regression (PLSR) and Random Forest Regression (RFR), were applied to predict the heavy metal concentration (Ba, Ni, Cr, Cu) of this specific Technosol. The coefficient of determination (R2) shows for Ba and Ni values between 0.50 (RMSE: 9%) and 0.61 (RMSE: 6%) for the PLSR and between 0.84 (RMSE: 0.03%) and 0.91 (RMSE: 0.02%) for the RFR model. The results for Cu and Cr show values between 0.57 (RMSE: 17.9%) and 0.69 (RMSE: 15%) for the PLSR and 0.86 (0.12%) and 0.93 (0.01%) for the RFR model. The pre-processing method, which improve the robustness and performance of both models best, is multiplicative scatter correction followed by the standard normal variate for the first and second derivatives. Random Forest in a first approach seems to deliver better modeling performances. Still, the pronounced differences between PLSR and RFR fits indicate a strong dependence of the results on the respective modelling technique. This effect is subject to further investigation and will be addressed in the upcoming analysis steps.