Machine learning-based mapping of soil organic carbon in agricultural bare soils at the watershed scale using Sentinel-2 time series, soil moisture and laboratory spectra

Accurate spatial quantification of soil organic carbon (SOC) is essential for assessing soil degradation, supporting sustainable land management and carbon accounting frameworks. However, SOC prediction based solely on multispectral optical imagery generally remains less accurate than approaches based on laboratory spectroscopy. This study aimed to map SOC content at the scale of the Naizin observatory watershed (12 km², Western France) by combining Sentinel-2 (S2), surface soil moisture (SM) time series and laboratory-derived spectral information and to assess the effect of the additional covariates on both model performance and the spatial distribution of SOC. Seven modeling approaches were tested to evaluate the contribution of different covariates using two machine learning algorithms, Random Forest (RF) and Deep Neural Network (DNN). The baseline model (M1) used only S2 bands as predictors, while subsequent models progressively incorporated S2 spectral indices, lab-converted S2 indices (LC-S2 indices) showing the strongest correlations with SOC, and SM covariates. The addition of S2 indices and SM resulted in only marginal performance improvements, whereas the addition of LC-S2 indices enabled higher improvements. Differences in model performance were directly reflected in the spatial patterns of the produced SOC maps. Models with higher predictive performance also produced spatial distributions that differed from those obtained using S2 bands alone, highlighting the sensitivity of spatial predictions to the selection of covariates. These results emphasize the importance of jointly evaluating statistical performance and spatial coherence when assessing machine learning models for SOC mapping.