Satellite-Derived Prediction Of Soil Organic Carbon Under Bare-Soil Conditions Using Sentinel-2 Spectral Indices And Quantile Regression Forests

Abstract

Soil Organic Carbon (SOC) is a primary indication of soil quality that impact nutrient availability, soil moisture, and carbon sequestration. Reliable SOC estimation is important for sustainable agricultural practices and to mitigate changes in climatic conditions. Conventional soil tests made in laboratory require more labor and they are spatially constrained. To overcome these limitation and challenges, this study proposes an operational framework for SOC prediction using multi-temporal Sentinel-2 Level-2A spectral bands and vegetation indices (VIs). The prediction models are built using the soil samples collected across three sampling periods—2017–2019, 2019–2021, and 2023–2024—in the Nilgiris, Erode, and Coimbatore districts of Tamil Nadu, India. For each soil sample the respective spectral and VI data were aggregated over the July–June interval over each sampling cycle. This helps to capture seasonal and inter-annual variations in vegetation and surface reflectance that influenced SOC accumulation. SOC prediction was performed using Quantile Random Forest for estimating both the central tendencies and uncertainty bounds. Model performance was evaluated using R², RMSE, MAE, and the Concordance Correlation Coefficient (CCC) and the SoC predictions are visualized through SOC distribution maps. The results demonstrate that integrating Sentinel-2 spectral bands, particularly red-edge and SWIR features, with different vegetation indices significantly improves prediction accuracy. The inclusion of VIs enhances the sensitivity of the model to vegetation vigor and organic matter inputs, providing a strong biophysical link to SOC variability. The proposed framework establishes a scalable, cost-effective, and replicable approach for regional scale SOC estimation using satellite remote sensing data and machine learning.