Integrating satellite-derived indices, edaphoclimatic variables and machine learning to predict soil organic carbon in northern Mexico rangelands

In recent years, the use of remote sensing and environmental data as proxies for soil organic carbon (SOC) has gained increasing attention; however, their effectiveness in heterogeneous dryland ecosystems remains uncertain. This study develops and evaluates an indicator-based modeling framework that integrates multispectral satellite indices, synthetic aperture radar (SAR) metrics, and edaphoclimatic variables to predict SOC in grassland and shrubland ecosystems of northern Mexico. A total of 276 composite soil samples (0–30 cm) were collected across 16 rangelands and linked to satellite-derived and environmental predictors. The predictive performance of three modeling approaches was compared, Random Forest, Extreme Gradient Boosting, and Multiple Linear Regression. Among them, Random Forest achieved the highest accuracy (R 2 = 0.80) and lowest prediction error (RMSE = 7.0 t ha −1 ). Variable importance analysis identified the Normalized Difference Lignin Index (NDLI), mean annual temperature, precipitation seasonality, and solar radiation as key indicators of SOC spatial variability. The resulting model enabled the generation of a spatially explicit SOC map, estimating a total stock of approximately 0.547 Gt C across grasslands and shrublands of Chihuahua. These results demonstrate that the integration of multi-sensor satellite-derived indicators and edaphoclimatic variables provides a robust and scalable approach for SOC prediction in dryland ecosystems. The proposed framework offers strong potential for application in data-scarce regions and supports the development of cost-effective tools for carbon monitoring, land management, and climate mitigation strategies.