In the framework of carbon farming, remote sensing (RS) is used to produce soil property maps for the upscaling of soil carbon stocking potentials, as modeled in in-field simulations. Reducing the uncertainty of the soil maps produced is a key element to obtain reliable upscaling products and finally calculate credible carbon credit potentials of the interested territories. The MRV4SOC project have tested the in-situ simulation of a RothC model in 3 demonstration sites in central-northern Italy, to calculate carbon credits of croplands and agroforestry. The soil input parameters used were soil organic carbon stock of the first 30 cm depth, and soil textural fractions, clay and sand. The MRV4SOC project has tested different procedures to produce soil map input parameters by remote sensing, such as: 1) a calibration of the bare soil composite of Europe SoilSuite (SRC), 2) a calibration of Sentinel and ENMAP spectral libraries using local field and lab-collected soil spectral libraries with VNIR spectroradiometers, and 3) a calibration of the SRC using a data augmentation strategy. AI models used for each calibration cases were: for case 1) quantile random forest (QRF) models, for case 2) XGBoost models, and for case 3) QRF and convolutional neural networks (CNN). The three strategies showed high calibration performance, but validation performance differs based on the model and strategy used, as well as the uncertainty of products.

Authors: Carlos Lozano Fondon, carlos.lozanofondon@crea.gov.it; Kevin Küelh, kevin.kuehl@dlr.de; Dimitra Palanza, dpalant@auth.gr; Konstantinos Karyotis, kkaryiotis@auth.gr; Costanza Andrenelli, mariacostanza.andrenelli@crea.gov.it; Roberto Barbetti, roberto.barbetti@crea.gov.it; Maria Fantappiè, maria.fantappie@crea.gov.it.