Towards Foundation Models for Zero-Shot Semantic Tree Segmentation

Kurzfassung

Tree detection and segmentation from remote sensing (RS) images is important for both forest management and urban planning. Previous methods using spectral sensors or height information to sepa rate trees from background are often costly. Supervised tree classification models are restricted in cross-domain generalization. While foundation models (FMs) have demonstrated strong zero-shot capabilities in natural image understanding and RS, their robustness in semantic tree segmentation in RGB imagery remains largely unexplored. We evaluate recent vision and vision-language FMs on three heterogeneous RGB datasets to assess their zero-shot generalization across unseen domains. GeoVLM achieves the most robust overall performance with a mIoU of 42.9 %, whereas SAM3 performs particularly well on high-resolution single-tree segmentation. The results demonstrate that geospatial vision-language models (VLMs) can serve for semantic tree segmentation in previously unseen domains, enabling downstream ecological and forestry analyses without requiring task-specific annotated training data.