Open-SAT: LLM-Enhanced Query Embedding for Satellite Image Retrieval

Researchers propose Open-SAT, a training-free algorithm that refines query embeddings using large language models (LLMs) to improve open-vocabulary object retrieval in satellite imagery. The method addresses the challenge of aligning natural language queries with satellite images, where vision-language models like CLIP often struggle. Open-SAT operates at inference time, leveraging LLMs to refine text embeddings and a vector database for efficient retrieval. The approach does not require additional training, making it practical for real-world applications. The paper is available on arXiv under ID 2605.05344.