Comparison of multivariate methods for inferential modeling of soil carbon using visible/near-infrared spectra

Abstract

In order to reduce costs and time in the analysis of soil properties, visible/near-infrared diffuse reflectance spectroscopy (VNIRS) has been proposed. Since various pre-processing transformations and calibration techniques are in use to analyze soil spectral data, much uncertainty still exists about predictive soil modeling. We investigated the feasibility of VNIRS to determine the concentration of carbon in soils collected in the Santa Fe River Watershed, Florida. A total of 554 soil samples (400 for calibration, and 154 for validation) were collected to a depth from 0 to 180 cm. Total carbon was measured by dry combustion, after sieving (2 mm), air-drying and ball-milling, and is reported in mg kg − 1 . Reflectance measurements from 350 nm to 2500 nm were collected in a controlled laboratory environment. Five multivariate techniques (stepwise multiple linear regression, principal components regression, partial least-squares regression, regression tree and committee trees) and thirty pre-processing transformations (including derivatives, normalization and non-linear transformations) of spectral data were compared with the aim of identifying the best combination to predict soil carbon. The coefficient of determination ( R 2), the root mean square error (RMSE), and the residual prediction deviation (RPD) were used to evaluate the models. The combination of multivariate technique and pre-processing transformation that provided the highest coefficient of determination for the validation set ( R v 2) and RPD, and lowest root mean square error for the validation set (RMSE v), was committee trees associated with Norris gap derivative with a search window of 7 measurements ( R v 2 = 0.86, RMSE v = 0.170, RPD = 2.68). When considering the overall results of the multivariate techniques across all tested pre-processing transformations, partial least-squares regression performed best (lowest average RMSE v across all pre-processing transformations), followed by stepwise multiple linear regression, and committee trees. In terms of pre-processing transformations, Savitzky–Golay derivatives consistently improved the models of soil carbon, being among the five best pre-processing transformations for all of the multivariate techniques tested. Norris gap derivative was the preferred data preparation for the tree-based techniques. Except for standard variate transformation, normalization techniques performed worse than expected. The RPD of the best VNIRS models were higher than 2.50, which suggest that the VNIRS models produced in this study are robust and stable enough to be applied for similar soils.