Randomized trial investigates LLM-generated prompts for improving zero-shot scene classification in remote-sensing imagery, suggesting potential advancements in model efficiency.
Key Points
This research aims to evaluate the effectiveness of meta-prompting with open-source language models for zero-shot scene classification in remote-sensing imagery.
Investigated three open-source LLMs: Mixtral-8×7B, Qwen 2.5 7B, and LLaMA 3.1 8B.
Utilized five remote-sensing benchmark datasets for testing.
Compared LLM-generated prompts to generic and handcrafted domain-specific prompts using various vision-language models.
LLM-generated prompts were competitive with manually designed templates in zero-shot scene classification.
In some cases, LLM prompts improved classification accuracy, depending on dataset and visual backbone.
The potential of open-source LLMs was reinforced as scalable prompt generators for remote-sensing recognition.