Improving soil organic carbon predictions from a Sentinel–2 soil composite by assessing surface conditions and uncertainties

• Sentinel-2 composites for large scale mapping of soil organic carbon in croplands. Identification pure pixels and quantification uncertainty. Normalized Burn Ratio < 0.05 removes crop residues moisture. Soil property improves when more than 6 scenes per pixel are used composite. Uncertainty reaches 5.5 g C kg –1. (SOC) prediction from remote sensing is often hindered by disturbing factors at the surface, such as photosynthetic active non–photosynthetic vegetation, variation moisture or surface roughness. With increasing amount freely available satellite data, recent studies have focused on stabilizing reflectance building using time series images. Although composite imagery has demonstrated its potential SOC prediction, it still not well established if resulting spectra mirror fingerprint optimal conditions to predict topsoil properties (i.e. a smooth, dry bare soil). We collected 303 photos surfaces Belgian loam belt where five main classes were distinguished: smooth seeded soils, crusts, partial cover growing crop, moist soils residue cover. Reflectance then extracted Sentinel–2 images coinciding with date photos. After was removed an NDVI 0.25, (NBR2) calculated characterize threshold NBR2 found be able separate unfavorable i.e. wet covered residues. Additionally, we that normalizing dividing each band mean all spectral bands) allows cancelling albedo shift between crusts seed–bed conditions. built exposed southern Belgium part Noord-Holland Flevoland Netherlands (covering spring periods 2016–2021). content means Partial Least Squares Regression Model (PLSR) 10–fold cross–validation. The uncertainty models assessed via interval ratio (PIR). cross validation model gave satisfactory results (mean 100 bootstraps: efficiency coefficient (MEC) = 0.48 ± 0.07, RMSE 3.5 0.3 –1 , RPD 1.4 0.1 RPIQ 1.9 0.3). maps show decreases number increases, minimum least six PIR 12.4 while predicted 14.1 ). against independent data set showed median difference 0.5 2.8 measured (average 13.5 ) contents field scale. Overall, this compositing method shows both realistic within regional patterns.