This study aims to predict carbon stocks for the state of Minas Gerais, Brazil, using machine learning (Random Forest Regressor) and a set of 41 environmental variables. The objective was to explore the model's dimensionality reduction to map carbon. For variable selection, the Altmann permutation method and top-k selection test were applied. Only solar radiation (Srad), water deficit (Def), and the Palmer Drought Severity Index (Pdsi) had significant effects for the final model. The results demonstrated that the simplified model presented slightly lower precision metrics than the complete model. As a result, the final model exhibits greater interpretability and robustness. Furthermore, non-linear relationships were found between carbon stock, solar radiation, and water stress. It is concluded that climatic factors are key factors influencing carbon distribution in the state. Optimised models are tools that can be utilised in public policies for conservation and management.