Insights into the Effects of Study Area Size and Soil Sampling Density in the Prediction of Soil Organic Carbon by Vis-NIR Diffuse Reflectance Spectroscopy in Two Forest Areas

Abstract

Sustainable forest land management requires measuring and monitoring soil organic carbon. Visible and near-infrared diffuse reflectance spectroscopy (Vis-NIR, 350–2500 nm), although it has become an important method for predicting soil organic carbon (SOC), requires further studies and methods of analysis to realize its full potential. This study aimed to determine if the size of the study area and soil sampling density may affect the performance of Vis-NIR diffuse reflectance spectroscopy in the prediction of soil organic carbon. Two forest sites in the Calabria region (southern Italy), which differ in terms of area and soil sampling density, were used. The first one was Bonis catchment area (139 ha) with a cover consisting mainly of Calabrian pine, while the second was Mongiana forest area (33.2 ha) within the “Marchesale” Biogenetic Nature Reserve, which is covered by beech. The two study areas are relatively homogeneous regarding parent material and soil type, while they have very different soil sampling density. In particular, Bonis catchment has a lower sampling density (135 samples out of 139 ha) than Mongiana area (231 samples out of 33.2 ha). Three multivariate calibration methods (principal component regression (PCR), partial least square regression (PLSR), and support vector machine regression (SVMR)) were combined with different pretreatment techniques of diffuse reflectance spectra (absorbance, ABS, standard normal variate, SNV, and Savitzky–Golay filtering with first derivative (SG 1st D). All soil samples (0–20 cm) were analyzed in the laboratory for SOC concentration and for measurements of diffuse reflectance spectra in the Vis-NIR region. The set of samples from each study area was randomly divided into a calibration set (70%) and a validation set (30%). The assessment of the goodness for the different calibration models and the following SOC predictions using the validation sets was based on three parameters: the coefficient of determination (R2), the root mean square error (RMSE), and the interquartile range (RPIQ). The results showed that for the two study areas, different levels of goodness of the prediction models depended both on the type of pretreatment and the multivariate method used. Overall, the prediction models obtained with PLSR and SVMR performed better than those of PCR. The best performance was obtained with the SVMR method combined with ABS + SNV + SG 1st D pretreatment (R2 ≥ 0.77 and RPIQ > 2.30). However, there is no result that can absolutely provide definitive indications of either the effects of the study area size and soil sampling density in the prediction of SOC by vis-NIR spectroscopy, but this study fostered the need for future investigations in areas and datasets of different sizes from those in this study and including also different soil landscapes.