Visual Prompting for Geospatial Image Segmentation based on Embedding Maps

The application of Foundation Models to Earth Observation (EO) data is often slowed down by complex data preparation for fine-tuning, creating the need for more efficient adaptation strategies.

This webinar features the Visual Prompting for Geospatial Image Segmentation (VP‑GIS), a framework that utilises high-dimensional embedding maps to guide zero-shot segmentation models toward domain-specific features. Rather than modifying model weights, the VP‑GIS approach injects learnable visual prompts into the input space, derived from the topological structures of pre-computed embedding manifolds. By aligning the visual prompts with the latent distribution of geospatial features (e.g., cosine similarity), the segmentation mask is created based on similarity scores between prompt and image embeddings.

The webinar showcases a working prototype demonstration, explains its functionality and opens the discussion on future work.