COMPARISON OF ARTIFICIAL NEURAL NETWORK (ANN) AND PARTIAL LEAST-SQUARES REGRESSION (PLSR) FOR ESTIMATING SOIL ORGANIC CARBON BY VIS–NIR SPECTROSCOPY: A CASE STUDY IN A KIWI ORCHARD IN SOUTHERN ITALY

Soil organic carbon (SOC) affects soil properties and processes as well as is an important sink and source of plant and microbial nutrients. Visible and near-infrared reflectance spectroscopy (Vis-NIR, 350–2500 nm) has become an important method for measuring and monitoring SOC. However, analysis of soil diffuse reflectance spectra data requires complex mathematical treatments to extract useful information and relate traditional measures of SOC concentrations to spectral data. Defining the most suitable methods for relating SOC measures to diffuse reflectance spectra is an active field of research which requires further study. The study was developed within E-Crops Project – Technology for Sustainable Digital Agriculture (National Operational Programme on R&I 2014–2020, Italy), and aimed to compare ANN and PLSR for estimating SOC by Vis–NIR spectroscopy. Topsoil (0-0.20 m) samples were collected at 126 locations within a kiwi orchard in Basilicata region (southern Italy) and analyzed for SOC concentration using a Shimadzu TOC-L analyzer. Vis-NIR reflectance spectra were measured in laboratory using an ASD FieldSpec IV 350– 2500 nm spectroradiometer. Accuracy and robustness of ANN and PLSR models were evaluated by a ten-fold cross-validation method based on R2, root mean square error (RMSE), and relative percent deviation (RPD). The dataset was randomly divided into ten equal sized groups (folds). One of the groups (10% of the dataset) was retained as validation data for testing the models, whereas the remaining groups (9 folds or about 90% of the selected data set) were used as training data. The process was repeated ten times, and each fold was used only once as validation data. The average of ten repeated processes was calculated to generate the SOC prediction models. The results showed that the prediction models obtained by ANN (mean R2= 0.94; mean RPD=4.07) performed better than those by PLSR (mean R2= 0.79; mean RPD=2.16). However, both PLSR model and ANN model were quite stable when calibration and validation sets were changed through the ten-fold cross-validation.