Optimizing Sentinel-2 temporal composites for soil organic carbon mapping and cropland management insights

Abstract

Accurate mapping of soil organic carbon (SOC) in intensive croplands is important for climate change mitigation and for guiding sustainable agricultural management. Despite the growing use of Sentinel-2 composites, evidence remains limited on how composite design affects SOC mapping accuracy in croplands and on whether satellite observations can capture management-relevant signals linked to SOC. This study compared four temporal Sentinel-2 spectral composites for SOC mapping using LUCAS 2015 and 2018 observations in Italy’s Po Plain. Three machine learning models, random forest, XGBoost, and CatBoost, were trained, and SHAP was used to interpret variable contributions. Across models, composites targeting the bare soil period, based on multispectral reflectance and non-photosynthetic vegetation indices, achieved the best performance. CatBoost performed best and produced a high-resolution SOC map for the Po Plain. In contrast, traditional vegetation indices such as NDVI and EVI showed limited relevance across all composites. Importantly, we found a robust negative association between SOC and bare soil frequency derived from multi-temporal Sentinel-2 observations, with lower bare soil frequency corresponding to higher SOC. This highlights bare soil exposure duration as a practical indicator for monitoring and suggests that management practices that shorten bare soil windows may help maintain or enhance SOC. Overall, this study optimized Sentinel-2 temporal composites with machine learning to improve SOC mapping in the Po Plain and provides actionable insights for cropland management in intensively cultivated regions.