Monitoring soil lead and zinc contents via combination of spectroscopy with extreme learning machine and other data mining methods

In order to limit pollution risk and develop proper remediation strategies, soil quality has to be controlled by rapid and sustainable monitoring measures. Visible and near-infrared reflectance spectroscopy (VisNIR) is an attractive surrogate to time-consuming and costly classical soil assessment protocols. It highly depends on selecting appropriate data mining methods for regression analysis. In this study, performance of a state of the art learning algorithm called extreme learning machine (ELM), was evaluated through comparing with the other calibration methods proposed in the literature for predicting lead (Pb) and Zinc (Zn) concentrations. Solid samples collected from a mine waste dump (n = 120) were scanned using a Fieldspec3 portable spectro-radiometer with a measurement range of (350-2500 nm) in a laboratory. Transformation of the reflectance spectra to absorbance was followed by three pre-processing scenarios including Savitzky-Golay smoothing (SG), first derivative (FD) and second derivative (SD). Partial Least Square Regression (PLSR), Support Vector Machine (SVM) and neural networks with two learning algorithms models (back propagation and extreme learning machine), were calibrated on spectral features selected by genetic algorithm, and then applied to predict soil metal concentrations. The best prediction accuracy was obtained by FD-ELM method with R(2)p, RMSEp, concordance correlation coefficient and RPD values of 0.93, 63.01, 0.98 and 5.92 for Pb and 0.87, 167.90, 0.91 and 5.62 for Zn, respectively. Study of the prediction mechanism proved that element sorption by spectrally active Fe-oxide and clay contents of the soil was the major mechanism by which the spectrally featureless Pb and Zn ions can be predicted. The spatial patterns of predicted toxic elements showed that FD-ELM had the most similarity with those maps obtained by interpolating measured values. Over all, it is concluded that reflectance spectroscopy combined with the ELM algorithm is a rapid, inexpensive and accurate tool for indirect evaluation of Pb and Zn and mapping their spatial distribution in dumpsite soils of Sarcheshmeh copper mine.