In-Context Learning for Few-Shot Change Detection in Satellite Imagery

Current Earth Observation (EO) change detection methods lack the flexibility to adapt to userdefined semantic queries. This project will develop a novel Vision-Language Model (VLM) that performs few-shot, in-context change detection. The model will interpret a natural language description and 0-8 visual examples to segment specific changes in satellite imagery, outputting precise polygons in a structured JSON format for direct GIS integration. We will: create a new fewshot benchmark; fine-tune a state-of-the-art VLM; advance in-context learning methods; integrate a multispectral foundation model for cross-satellite capability; and build a demonstrator web application.