Urban tree vegetation plays a key role in sustainable urban planning and ecosystem service provision. This study evaluates the performance of the Segment Anything Model (SAM), developed by Meta AI, in the segmentation of urban tree vegetation from orbital PlanetScope imagery. These images were selected due to their high spatial and temporal resolution, which makes them particularly suitable for urban applications. SAM was applied in zero-shot mode, guided by geometric prompts over representative tree-covered areas. The analysis was conducted across three Brazilian cities—Corumbá (MS), Rio Verde (GO), and Valparaíso de Goiás (GO)—using different spectral band compositions. SAM’s performance was evaluated through a combined quantitative and qualitative approach, using reference masks derived from manually annotated tree canopy polygons. Although SAM had not been previously trained on satellite imagery, it achieved an F1-scores close to 70% and recall values around 75%, independently of the spectral band composition provided as input. These results demonstrate the model’s generalization ability—even under spectrally constrained scenarios involving only three bands. Qualitative analysis confirmed spatial consistency in tree crown delineation, particularly in homogeneous areas, while over-segmentation was observed in spectrally heterogeneous environments. While the results are promising for exploratory and semi-automated vegetation mapping, they also underscore need for fine-tuning SAM on satellite data to enhance spatial precision and thematic discrimination. Overall, SAM's modular and prompt-based architecture offers a robust foundation for scalable, supervised remote sensing workflows focused on urban vegetation monitoring.