Estimating soil heavy metals concentration at large scale using visible and near-infrared reflectance spectroscopy

This study was aimed (i) to examine using diffuse reflectance spectra within VNIR region to estimate soil heavy metals concentrations at large scale, (ii) to compare the influence of different pre-processing models on predictive model accuracy, and (iii) to explore the best predictive models. A number of 325 topsoil samples were collected and their spectral data, pH, clay content, organic matter, Ni, and Cu concentrations were determined. To improve spectral data, various pre-processing methods including Savitzky-Golay smoothing filter, Savitzky-Golay smoothing filter with first and second derivatives, and standard normal variant (SNV) were used. Partial least squares regression (PLSR), principal component regression (PCR), and support vector machine regression (SVMR) models were employed to build calibration models for estimating soil heavy metals concentration followed by evaluation of provided predictive models. Results indicated that Cu had stronger correlation coefficients with spectral bands compared to Ni. Cu and Ni demonstrated strongest correlations at wavelengths 1925 and 1393 nm, respectively. Based on RMSE, R², and RPD statistics, the PLSR model with Savitzky-Golay filter pretreatment provided the most accurate predictions for both Cu and Ni (R² = 0.905, RMSE = 0.00123, RPD = 2.80 for Ni; R² = 0.825, RMSE = 0.00467, RPD = 2.04 for Cu) where such prediction was much better for Ni than for Cu. Reasonable results with lower accuracy and stability were obtained for PCR (R² = 0.742, RMSE = 0.00181, RPD = 1.91 for Ni; R² = 0.731, RMSE = 0.00578, RPD = 1.65 for Cu) and SVMR (R² = 0.643, RMSE = 0.00091, RPD = 3.80 for Ni; R² = 0.505, RMSE = 0.00296, RPD = 3.22 for Cu). We concluded that reflectance spectroscopy technique could be applied as a reliable tool for detection and prediction of soil heavy metals.