Reproduce “Few-Shot Remote Sensing Image Scene Classification with CLIP and Prompt Learning” on OpenResearch

About this paper

Ivica Dimitrovski, Vlatko Spasev, Ivan Kitanovski

Researchers at the University Ss Cyril and Methodius explored prompt learning strategies to adapt Vision-Language Models like CLIP for few-shot remote sensing image scene classification. Their systematic evaluation across nine datasets showed that prompt-learning methods, particularly PromptSRC and MaPLe, consistently achieved higher classification accuracies than zero-shot CLIP and linear probe baselines, with PromptSRC demonstrating superior cross-dataset generalization.

Key insights

• Prompt learning effectively bridges the domain gap between natural and remote sensing imagery, allowing powerful VLMs to adapt efficiently with minimal labeled data.
• Multi-modal prompt learning (MaPLe) and methods incorporating self-regulating constraints (PromptSRC) significantly enhance performance by improving cross-modal alignment and preventing catastrophic forgetting.
• Prompt-learning approaches can learn highly transferable representations that generalize well across diverse remote sensing datasets, even with substantial domain shifts.

Summary by alphaXiv.