Influence of the calibration set size on the prediction of soil organic carbon through Vis-NIR spectroscopy

The aim of the work was to evaluate the effect of the calibration set size on the predictive performance of Vis-NIR models for soil organic carbon (SOC).The dataset is made up of 217 topsoils (0-20 cm) collected in the "Marchesale" Biogenetic Nature Reserve (Serre Massif, Calabria) within the project LIFE09 ENV/IT/078 "Managing forests for multiple purposes: carbon, biodiversity and socio-economic wellbeing" (ManFor C.BD.). Soils, developed in a beech forest onto Paleozoic granitoid rocks or colluvial deposits, are young (Entisol and Inceptisol) and shallow to moderately deep. The samples, oven dried at 40° for 48 hours and sieved at 2 mm, were used for spectroscopic measurements and analysed for SOC content. The 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), whereas SOC was determined using a TOC-analyzer (Shimadzu Corporation, Kyoto, Japan). Reflectance was converted to absorbance and a standard normal variate (SNV) pre-processing was applied to each spectrum.Principal component regression (PCR) and support vector machine regression (SVMR) were used to analyze the relationships between spectra and SOC whereas root mean square error (RMSE) and coefficients of determination (R2) were used to evaluate the sensitivity of the models. Calibration data, from 45 to 80% of total samples, were selected within the deciles of the sample distribution, and used to construct the regression models, validated by means of independent data.Results revealed that when the number of calibration set samples was relatively small, PCR generated models with poor generalization ability. An increasing of R2 and low RMSEs were observed until the calibration set samples were about 70% of the total. On the other hand, when the SVMR was used, the sampling size was not a critical issue.