Hybrid machine learning-process-based models for soil organic carbon prediction in drylands: A systematic review

Abstract

CONTEXT Dryland ecosystems store a substantial proportion of global soil organic carbon (SOC) and contribute to climate regulation and land sustainability. However, SOC dynamics in drylands are highly complex because they are influenced by pulsed rainfall, microbial dormancy, photodegradation, erosion, and increasing climate variability under future warming scenarios. Traditional process-based models provide mechanistic understanding of carbon cycling but are constrained by high data requirements, parameter uncertainty and limited scalability. In contrast, machine learning (ML) models offer strong predictive capability and efficient spatial scaling using remote sensing and environmental datasets, although they often lack mechanistic interpretability and robustness under changing climate conditions. OBJECTIVE This study synthesised advances in hybrid ML-process-based approaches for dryland SOC prediction between 2015 and 2026. METHODS Peer-reviewed studies published between 2015 and 2026 on dryland SOC modelling were systematically retrieved and screened following the PRISMA framework. The selected studies were synthesised to evaluate major biophysical controls on dryland SOC, compare process-based and AI-driven models, and develop a taxonomy of hybrid architectures, including process-guided ML, physics-informed neural networks, and ensemble data-fusion systems. RESULTS AND CONCLUSIONS Earth observation technologies such as Landsat, Sentinel, and MODIS are critical for large-scale SOC monitoring and carbon accounting. Although hybrid frameworks significantly improve predictive accuracy and scalability, major challenges remain regarding uncertainty quantification, long-term validation, sparse field observations, and model interpretability. SIGNIFICANCE This review signifies hybrid mechanistic-AI frameworks represent a promising pathway for improving dryland SOC prediction and supporting climate mitigation, land restoration, and sustainable ecosystem management.