Abstract

Soil organic carbon (SOC) represents a critical component of global carbon cycling and serves as a fundamental indicator of soil health and agricultural sustainability. This study integrates remote sensing data with machine learning algorithms to quantify SOC changes in agricultural croplands under varying climatic conditions. We employed multi-temporal Landsat 8 OLI and Sentinel-2 MSI imagery combined with climate variables to develop predictive models for SOC estimation across diverse cropping systems. Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN) algorithms were evaluated using field-collected SOC measurements from 450 sampling sites across three distinct agro-climatic zones over a five-year period (2018-2022). The RF model demonstrated superior performance with R² = 0.78 and RMSE = 2.34 g kg⁻¹, followed by ANN (R² = 0.75, RMSE = 2.58 g kg⁻¹) and SVM (R² = 0.69, RMSE = 2.91 g kg⁻¹). Spectral vegetation indices, particularly the Normalized Difference Vegetation Index (NDVI) and Soil Adjusted Vegetation Index (SAVI), showed strong correlations with SOC content (r > 0.65). Climate variables including temperature and precipitation patterns significantly influenced SOC dynamics, with temperature showing negative correlations (-0.58) and precipitation showing positive correlations (0.43) with SOC accumulation. The integrated approach successfully mapped SOC changes at 30-meter spatial resolution, revealing annual SOC loss rates ranging from 0.2-0.8% across different land management practices. These findings provide valuable insights for precision agriculture applications and carbon sequestration monitoring in agricultural landscapes.