Carbon stock modelling in Minas Gerais, Brazil: effects of dimensionality reduction in machine learning algorithms

This study aimed to forecast soil carbon in Minas Gerais, Brazil. We are using the Random Forest Regressor machine learning technique. The study utilised 41 environmental factors. We were focused on how reducing the model’s complexity could help map soil carbon more effectively. To select the variables, we utilised Altmann’s permutation method combined with the top-k selection test. Results show that solar radiation, water deficiency, and the Palmer Drought Severity Index were the main variables. Notably, although the simplified model exhibited precision metrics, it provided greater interpretability and robustness. We also found several non-linear connections between carbon storage and variables such as solar radiation and water stress. It became evident that the climate significantly influences carbon distribution in this area. In general, these enhanced models could significantly support agricultural zoning, conservation initiatives, and soil management.