Modeling soil organic carbon in the Brazilian amazon with geostatistical and machine learning techniques

• Random Forest achieved the greatest SOC content accuracy, while IDW best predicted SOC stock. • SOC content and stocks consistently followed forest > pasture > cropland across the Brazilian Amazon region. • Forest soils had the highest predicted SOC, with 6.59% at 30–60 cm and 261.95 Mg C ha⁻¹ at 0–30 cm. • Clay content, temperature, and land use were the primary environmental drivers of SOC variability. Reliable estimation of soil carbon storage is essential to quantify human effects on the terrestrial carbon cycle. In this work, we analyzed soil organic carbon (SOC) content and stocks in the Brazilian Amazon at two depths (0–30 and 30–60 cm) using 486 georeferenced samples that were evaluated for texture, pH, SOC, and bulk density. In this study, inverse distance weighting (IDW) and machine learning algorithms (Random Forest, Support Vector Machine, Multiple Linear Regression, and Artificial Neural Network) were applied and validated through 10-fold cross-validation. Random Forest had the highest accuracy for SOC content (R² = 0.96), and IDW was the best method for SOC stock (R² = 0.88). Predicted SOC content and stock also followed the trend forest > pasture > cropland, and varied from 2.71–4.52 % (103.89–261.95 Mg C ha−¹) in forest soils to 2.03–2.14% (80.05–86.06 Mg C ha −1 ) on croplands. Clay content, temperature, and land use were the major factors influencing SOC variation. It also indicates the possibility of combining IDW and machine learning approaches to bridge data gaps to guide climate-resilient land management and improvement of SOC, not only in the Amazon but across other tropical and subtropical areas where soil data is sparse. By delivering scalable and transferable modeling systems, this work enables a wider scope for impact in global carbon accounting, sustainable agriculture, and land-use planning for climate mitigation.