Prediction of Soil Properties by Visible and Near-Infrared Reflectance Spectroscopy

Abstract

The major goal of this research was to assess the ability of visible-near-infrared (VIS-NIR) spectroscopy combined with multivariate regressions based on four modeling algorithms, namely partial least squares regression (PLSR), principal component regression (PCR), support vector regression(SVR), and random forest (RF) to quantify some quality soil parameters. Saturated electrical conductivity (ECe), organic carbon (OC), and texture (including sand, silt, and clay) were measured. Overall, 114 soil samples were taken from 30 profiles (0–100 cm) in Doviraj plain, western Iran. For improving spectral data, various preprocessing techniques including Savitzky-Golay first deviation (SGD), Normalization (N), and Standard Normal Variate transformation (SNV) were employed. A leave-one-out cross-validation procedure was used. Results revealed that RF model produced much more precise predictions compared with PCR, SVR, and PLSR models and specifying the most important variables for the model. Also, the highest accuracy for RF model was achieved by SGD preprocessing for ECe, OC, sand, and clay with validation coefficient of determination (R2val) equal to 0.57, 0.75, 0.70, and 0.73, respectively; also, validation root mean square error (RMSEval) was equal to 0.74 dS/m, 0.13, 12.97, and 0.14%, respectively, However, SNV preprocessing with RF model caused to best result with R2val = 0.65 and RMSE val = 9.21% for silt. This methodology displays the capability of the VIS-NIR for estimation of ECe, OC, and texture.