Effect of calibration set size on prediction at local scale of soil carbon by Vis-NIR spectroscopy

Predicting soil properties through visible and near-infrared (Vis-NIR) spectroscopy by a limited number of calibrationsamples can reduce the cost and time for physic-chemical analyses. This study was aimed to assess theinfluence of calibration set size on the prediction of total carbon (TC) in the soil by Vis-NIR spectroscopy. In a forestedarea of 33 ha in southern Italy (Calabria), 216 soil samples were analyzed for TC concentration, and reflectancespectra were measured in the laboratory. The whole data set was randomly split into calibration andvalidation sets (70% and 30%, respectively). To study the effect of the number of samples on TC prediction, tencalibration subsets of samples between 14 and 144 were selected. Three techniques including principal componentsregression (PCR), partial least squares regression (PLSR) and support vector machine regression (SVMR)were used to develop 84 calibration models, validated through the same independent data. The models werecompared through the coefficient of determination (R2), the root mean square error of prediction (RMSEP)and the ratio of the interquartile distance (RPIQ). Validation results showed that to obtain not significant differenceswith models based on the full calibration set, 29, 72 and 115 samples were required for PCR, SVMR andPLSR respectively. Although PCR appeared less sensitive than PLSR and SVMR to calibration sample size, SVMRoutperformed PLSR and PCR with higher R2 and RPIQ values and lowerRMSEP. To obtainRMSEP not significantlydifferent fromthe best model achieved in this study, the required minimumnumber of sampleswas 72 for SVMRand 130 for PLSR. All PCR model were significantly poorest than the best model.