Contribution of Sentinel-2 Seedbed Spectra to the Digital Mapping of Soil Organic Carbon

Soil organic carbon (SOC) plays a pivotal role in the functioning of terrestrial ecosystems, has the potential to mitigate climate change and provides several benefits for soil health. Understanding the spatial distribution of SOC can possibly help formulate sustainable soil management practices. Conventional soil surveys are often limited by their spatio-temporal resolution and high cost, necessitating the development of innovative techniques that can capture the intricate variability of SOC across landscapes. In response to this need, digital soil mapping (DSM) has emerged as a powerful approach that uses advanced geospatial technologies and statistical methods to predict soil properties across large areas. Predictor variables for DSM include climate data, topographical features, geological attributes, legacy soil maps, land management practices, spatial information and remote sensing data. The spectral response of bare soil, measured by multispectral satellite sensors, can be an adequate predictor of SOC and texture at the field scale and in small regions, but its use for the assessment of soil properties at large scale (thousands of km²) has been less explored 1. In this study, bare soil spectra derived from Sentinel-2 were used to estimate SOC and texture across agricultural parcels in Flanders, northern Belgium (n=169-175). Five different machine learning models were tested: generalized linear regression (GLM), partial least squares regression (PLSR), random forest (RF), cubist regression (CR) and gradient boosting machine (GBM). The SOC prediction of a DSM model using bare soil spectra was compared with that of a DSM model using environmental covariates: topography (elevation, slope and compound topographic index), climate (average annual temperature, total annual precipitation, average annual evapotranspiration), texture (sand, silt and clay content), vegetation (proportion of the year the soil is covered by vegetation) and location. The predictive performances of these models were compared to a DSM model that included both the bare soil spectra and the environmental covariates. Soil texture (sand, silt, clay) was adequately predicted using the bare soil spectra from the spring seedbed (R²: 0.53-0.71; RPD: 1.54-2.18; RPIQ: 1.36-2.41), but the predictive performance for SOC was poor (R²: 0.19; RPD: 1.07; RPIQ: 1.45). All three DSM models showed poor predictive performance for SOC, with the best performance for the model including all covariates (R²: 0.26; RPD: 1.25; RPIQ: 1.68). For the DSM model from bare soil spectra (PLSR), all Sentinel-2 spectral bands showed high relative importance except bands 2 (blue) and 3 (green). For the DSM model from environmental covariates (GBM), vegetation cover and topography explained most of the variation in SOC. The DSM model including all variables (GBM) showed a low influence of bare soil spectral bands, but a high influence of previous vegetation cover and topography. These results showed the importance of terrain characteristics and vegetation for assessing large scale SOC distribution. The overall low predictive performance for SOC obtained in this study indicates the complex nature of factors influencing SOC distribution across a large region and highlights the need for more in-depth high resolution studies. 1 https://doi.org/10.3390/rs14122917