Beyond Templates: Revisiting Zero-Shot Remote Sensing through Meta-Prompting

The article presents a comprehensive evaluation of 17 vision-language model (VLM) variants applied to zero-shot Earth Observation tasks using Meta-Prompting for Visual Recognition (MPVR) across 12 remote sensing datasets. It highlights the sensitivity of zero-shot performance to the design of textual prompts and class descriptions, demonstrating that while LLM-generated descriptions are semantically richer, they can introduce noise that undermines robustness. The study emphasizes the effectiveness of lightweight query embedding calibration in enhancing zero-shot classification and retrieval, providing valuable insights for practitioners in optimizing model performance in remote sensing applications.