Identifying mining-induced chromium contamination in soil through visible-near infrared spectroscopy and machine learning

The extraction, purification, and utilization of mineral resources have been among the largest anthropogenic sources of chromium (Cr) in soil. Determining Cr contamination in soil is a key issue prior to its appropriate remediation. Nevertheless, the efficient identification of large-scale soil Cr contamination requires continuous research. The present study proposes a continental-scale method to rapidly identify soil Cr contamination using visible-near infrared spectroscopy (vis-NIR) and machine learning (ML). A large dataset containing 18,675 topsoil samples from the Land Use/Land Cover Area Frame Survey 2009 projects across Europe was compiled. Five advanced ML algorithms were compared, and hyperparameter optimization was conducted using the grid search method. Permutation importance was employed to calculate the rank of each spectral wavelength, shedding light on the most sensitive spectral wavelength for Cr contamination. Results indicate that hyperparameter optimization had the most significant performance improvement on support vector machine (SVM), exhibiting an increase in training performance from 0.795 to 0.868. The achieved optimal SVM accuracy, area under the receiver operating feature curve, sensitivity, and specificity of 0.78, 0.85, 0.85, and 0.66, respectively, indicating excellent predictive performance on the Cr contamination classification. The optimal SVM model revealed that the most important spectral band for classifying Cr contamination was 1430–1433 nm. This finding implies that the adsorption of molecular water was closely related to the classification of Cr contamination. The current study introduces the first continental-scale identification of Cr contamination using vis-NIR, which has excellent guiding significance for Cr remediation and the identification of other heavy metals using vis-NIR.