Advancing Soil Organic Carbon Prediction: A Comprehensive Review of Technologies, AI, Process-Based and Hybrid Modelling Approaches

Abstract

Measurement, monitoring, and prediction of soil organic carbon (SOC) are fundamental to supporting climate change mitigation efforts and promoting sustainable agricultural management practices. This review discusses recent advances in methodologies and technologies for SOC quantification, including remote sensing (RS), proximal soil sensing (PSS), artificial intelligence (AI) for SOC modelling (in particular, machine learning (ML) and deep learning (DL)), biogeochemical modelling, and data fusion. Integrating data from RS, PSS, and other sensors usually leads to good SOC predictions, provided it is supported by careful calibration, validation across diverse pedo-climatic and land management, and the use of data processing and modelling frameworks. We also found that the accuracy of AI-driven SOC prediction improves when RS covariates are included. Although DL often outperforms classical ML, there is no single best AI algorithm. By incorporating simulated outputs from biogeochemical model as additional training data for AI, causal relationships in SOC turnover can be incorporated into empirical modelling, while maintaining predictive accuracy. In conclusion, SOC prediction can be enhanced through 1) integrating sensing technologies, 2) applying AI, notably DL, 3) addressing biogeochemical model limitations (assumptions, parameterization, structure), 4) expanding SOC data availability, 5) improving mathematical representation of microbial influences on SOC, and 6) strengthening interdisciplinary cooperation between soil scientists and model developers.