Synergetic integration of multi-temporal remote sensing mosaic and conventional soil map for mapping organic carbon content in Chernozems

The spatial distribution of soil organic carbon (SOC) is assessed using digital soil mapping (DSM) methods, where remote sensing data serve as a key factor. Although soil maps provide valuable information, they are not frequently employed as predictors in soil modelling. This study explored and compared the utilisation of a multi-temporal mosaic of Sentinel-2 (S2) data, in combination with a conventional soil map, for SOC mapping in Chernozem soils of the forest-steppe zone. We implemented two algorithms to estimate SOC content in topsoil (0-30 cm) under several scenarios. As initial predictive covariates, we used a temporal mosaic of S2 bare soil spectra and a large-scale soil type map. We first implemented either partial least square regression (PLSR) or random forest (RF) to predict SOC from S2-derived bare soil spectral responses. Using solely the bare soil temporal mosaic spectra, PLSR proved more effective than RF to predict SOC. The PLSR-derived map of SOC predictions was then combined with the soil type map using RF. This latter combination led to the best SOC prediction performance and least uncertainty. Our findings indicate that integrating a soil map into a remote-sensing-based DSM prediction of SOC yields benefits in SOC mapping compared to using solely remote sensing or soil map data.