Bayesian geostatistical modeling outperforms machine learning for mapping carbon stocks from Sentinel-2 data in plantation-forest landscapes of Zimbabwe’s Eastern Highlands

Introduction Accurate, spatially explicit estimates of carbon stocks in plantation forests are essential for sustainable management and credible climate-change mitigation, yet plantation mosaics often exhibit strong spatial structure that can reduce the reliability of models that ignore spatial dependence. Methods We predicted aboveground carbon stock in planted-forest landscapes of Nyanga District, Eastern Highlands, Zimbabwe, using multispectral Sentinel-2 data and compared a Bayesian geostatistical hierarchical model with three machine-learning approaches: random forest, support vector machine, and k-nearest neighbors. Predictors included NDVI, SAVI, EVI, and distance to settlements, and models were evaluated using spatially blocked cross-validation. Results The spatially informed Bayesian geostatistical hierarchical model delivered the lowest prediction errors and outperformed the best machine-learning model, random forest, while support vector machine and k-nearest neighbors showed larger errors. NDVI was the strongest predictor, SAVI added complementary signal, EVI contributed limited additional skill, and distance to settlements captured anthropogenic influence near managed edges. Bayesian modeling also enabled probabilistic uncertainty mapping, with higher uncertainty in sparsely sampled or rapidly changing plantation patches. Discussion Bayesian spatial modeling provides a robust and uncertainty-aware framework for planted-forest carbon accounting and outperforms the compared machine-learning approaches in this heterogeneous plantation landscape. These findings support its use for monitoring, reporting and verification and for site-level management decisions aligned with climate-mitigation and ecosystem-service goals.