A hybrid machine learning framework for land use carbon accounting: A case study of Tanzania

This study presents a structured integration of land-use carbon accounting with regression, time-series, and machine-learning models to examine historical patterns and prospective trajectories of land-use related CO2 emissions in data-constrained settings. Rather than proposing a new accounting methodology, the framework demonstrates how established modeling approaches can be combined to support comparative analysis and scenario exploration. The study demonstrate the framework’s application through a case study of Tanzania, integrating multi-source land-use and socio-economic data within a hybrid ensemble of multiple linear regression, ARIMA time-series modeling, and machine learning approaches (Random Forest and XGBoost). The analysis indicates that land-use change explains a substantial share of modeled emissions variability within the accounting-consistent framework, with cropland-to-forest conversion associated with comparatively larger modeled emission reductions under the explored scenarios. These results reflect model-based associations rather than causal dominance and are conditional on the accounting structure and scenario assumptions. In contrast, Socio-economic drivers, particularly urbanization and economic growth, were associated with variations in modeled emissions, although the magnitude and direction of these effects depend on model specification. Scenario analysis suggests that a 20% conversion of cropland to forest is associated with a reduction in modeled emissions from 24,339–18,041 metric tons, whereas combined urbanization and GDP growth increase projected emissions. Machine-learning models, particularly XGBoost, exhibited lower prediction errors under the adopted validation design; however, because the data are temporally indexed, these results should be interpreted as measures of internal predictive consistency rather than strict out-of-sample forecasting accuracy. Overall, the framework provides evidence-informed and conditional insights that may support prioritization of land-based mitigation options in developing-country contexts under the Paris Agreement, while remaining contingent on model specification, data structure, and validation design.