Accurate carbon stock classification is vital for supporting climate change mitigation efforts. Traditional methods are expensive and time-consuming, prompting the adoption of remote sensing techniques combined with machine learning for efficiency. This study evaluates the performance of XGBoost and Random Forest classifiers using drone and Google Earth Engine (GEE) imagery, with VGG16 applied as a feature extractor. Data collected from field plots at Telkom University, Bandung, Indonesia, were labeled into three classes: low, medium, and high carbon stock. The drone dataset used in this experiment consists of 2,114 images, while the GEE dataset comprises 2,526 images. This experiment results demonstrate that XGBoost with drone imagery achieves the highest accuracy of 90.79%, outperforming Random Forest and GEE-based models.