Artificial intelligence for forest carbon dynamics and climate mitigation strategies

Forests are central to the global carbon cycle, yet reliable estimation of carbon stocks, fluxes, disturbances, and recovery remains difficult across heterogeneous landscapes. This study used a systematic and integrative review to synthesize empirical studies published between 2010 and July 2026 on artificial intelligence (AI), machine learning, and deep learning in forest-carbon assessment and climate-mitigation applications. Using a PRISMA-S approach about 95 articles were synthesised from a pool of 254 based on model families, data sources, validation designs, uncertainty reporting, transferability, and operational relevance. The review identified that AI has expanded capacity for aboveground-biomass estimation, carbon-stock mapping, carbon-flux modeling, deforestation and degradation monitoring, disturbance assessment, and digital monitoring, reporting, and verification (MRV). However, AI estimation performance was highly contingent on ecosystem, response range, spatial scale, reference-data quality, predictor composition, and validation design. Only 9% of studies fully and 5% partially addressed independent external validation of models used; the remaining did not. The overall appraisal classified 2% of studies as high confidence, 50% as moderate confidence, and 48% as limited confidence. Studies improved model performance and results through multi-source integration and local calibration producing more conservative estimates than random data splits or product-to-product comparisons. Deep-learning and stacked-ensemble approaches were most convincing in data-rich settings, but the evidence did not establish their universal superiority over Random Forest or boosting models. Most of the studies had persistent constraints such as sparse reference data in underrepresented regions, scaling uncertainty, limited interpretability, inconsistent reporting, and weak evidence of cross-region transfer. Credible AI-enabled forest-carbon systems therefore require transparent data partitioning, independent validation, explicit and scale-specific uncertainty analysis, reproducible workflows, and regional calibration.