Soil organic carbon prediction using Sentinel-2 data and the LUCAS topsoil database

Abstract: Its carbon sink potential as well as soil fertility benefits make organic carbon a soil variable for which reliable quantification methods are sought. This thesis work aims at investigating the possibility of adapting a large soil spectral library to build models for SOC predictions with remotely sensed, multispectral data. For this purpose, the continental-scale LUCAS topsoil database was spectrally resampled to simulate the reflectance measured by the Sentinel-2 satellite. Multivariate partial least squares regression models were created based on the spectrally resampled LUCAS database (i)for all mineral cropland soil samples and (ii) for a regional subset of the mineral cropland samples relative to location of the validation samples in southern Sweden. The global model was poor (RPD= 1.09) in relation to a comparable model produced with the original spectral information. This outcome was related to the insufficient spectral information of Sentinel-2 type data to account for the variability of soil chromophores within this large dataset. Despite the reduced extent and a samplesize (n = 70) that is comparable to moderately successful SOC modelling attempts, the regional model yielded only a slight performance improvement (RPD = 1.12). Reasons for this outcome could be the spatially dispersed sampling strategy used to collect the LUCAS database or the high sand content of the samples. Both models failed to produce reasonable predictions of the validation dataset. The investigation of the difference between spectrally resampled LUCAS reflectance and remotely sensed Sentinel-2 reflectance revealed that the data of these two measurement methods are not readily compatible.