Spatial variability of soil organic carbon in Tamil Nadu: Integrating remote sensing and machine learning approaches

Aim: The experiment was carried out to estimate the geographical distribution of soil organic carbon (SOC) in Tamil Nadu, India, using advanced machine learning models. Methodology: High-resolution remote sensing data from Landsat 8, climate indices from TerraClimate, and terrain indices from the Shuttle Radar Topography Mission (SRTM) were integrated to construct reliable SOC estimation models. The models employed in this study included Random Forest (RF), Quantile Neural Network (QNN), Cubist, and Bootstrapped Random Forest (BRF). These models were selected for their ability to capture complex relationships between SOC and environmental covariates. Results: The QNN and Cubist models displayed superior accuracy with mean absolute error (MAE) values of 0.48 and 0.47, respectively. Variogram analysis indicated moderate spatial dependence across all models, with the Cubist model exhibiting the highest partial sill and sill values. Interpretation: These findings demonstrate the effectiveness of combining machine learning with remote sensing for SOC estimation. The results suggest that these models can be valuable tools in improving the accuracy of SOC distribution maps, which are essential for sustainable crop production. By identifying potential SOC sequestration areas, this research contributes to the development of targeted soil management strategies that prioritize soil health and support sustainable agricultural practices.