Abstract

Soil carbon sequestration represents a critical nature-based solution for climate change mitigation, yet accurate quantification of carbon stocks and sequestration potential remains challenging across landscape scales. This study developed an artificial intelligence framework integrating deep learning algorithms with multi-source environmental data to predict soil organic carbon (SOC) stocks and identify areas with high sequestration potential. We analyzed 1,850 soil profiles across a 3,200 km² agricultural region, combining spectral data from Sentinel-2 and hyperspectral sensors, climate variables, topographic attributes, land management history, and soil physicochemical properties. A novel deep neural network architecture incorporating attention mechanisms achieved R² = 0.94 for SOC stock prediction (0-100 cm depth) with RMSE of 8.7 Mg C ha⁻¹. The AI model identified 42% of the study area with high sequestration potential (>20 Mg C ha⁻¹ additional storage capacity), primarily in degraded croplands and grasslands. Scenario modeling revealed that optimized management practices could sequester 2.3 Tg C over 20 years, equivalent to 8.4 Tg CO₂ removal. Feature importance analysis highlighted vegetation indices, clay content, and precipitation as key predictors. The framework provides spatially explicit guidance for carbon farming initiatives, supporting evidence-based policy development and verification of carbon credits. This research demonstrates AI's transformative potential for scaling soil carbon assessment and optimizing climate mitigation strategies.