Estimating soil organic carbon with soil Vis-NIR spectroscopy: a comparison of different algorithms in two forest areas

Sustainable forest land management requires the development of rapid, accurate, and cost effective methods to determine forest soil organic carbon (SOC). Reflectance spectroscopy in the visible, near infrared (Vis-NIR) region could be an alternative to standard laboratory methods. The work was aimed to evaluate the performance of different multivariate calibrations techniques analysing soil reflectance spectra to estimate soil organic carbon (SOC). The study was developed within two study areas located in southern Italy: (1) the Bonis catchment (139 ha), mainly covered by Calabrian pine, and (2) the "Marchesale" Biogenetic Nature Reserve (33.2 ha) covered by forest beech. The two study area show relatively homogeneous features in terms of parent material and soil type, whereas they differ in soil sampling density: 135 samples collected in the Bonis catchment and 231 in the "Marchesale" Biogenetic Nature Reserve. In both areas, soil samples were collected up to a depth of 0.20 m, oven dried at 40° and sieved at 2 mm, and then used for spectroscopic measurements and analyses of SOC content. SOC content was determined using a TOC-analyzer (Shimadzu Corporation, Kyoto, Japan), whereas Vis-NIR reflectance was measured in laboratory, under artificial light, using an ASD FieldSpec IV 350-2500 nm spectroradiometer (Analytical Spectral Devices Inc., Boulder, Colorado, USA).To reduce the amount of data and computation time, the spectra were averaged every 10 nm. Spectral reflectance (R) was transformed to apparent absorbance (A) by A = Log (1/R). Each data set was randomly split into a calibration and a validation set (70% and 30%, respectively). Three techniques including principal components regression (PCR), partial least squares regression (PLSR) and support vector machine regression (SVMR) were used to estimate the calibration models, which were validated through the independent datasets. The models were compared through the coefficient of determination (R2), root mean square error of prediction (RMSEP) and interquartile distance (RPIQ). For both study areas, the results showed that PLSR outperformed with higher R2 and RPIQ values and lower RMSEP. PCR showed similar results than PLSR but with lower R2 and RPIQ values and higher RMSEP, whereas the worst results were obtained for SWMR. These preliminary results bring out the need to analyse the effects of different factors (soil type, sample size, etc.) on the relative performance of the prediction techniques.