Boosting Few-Shot Remote Sensing Image Scene Classification with Language-Guided Multimodal Prompt Tuning

Remote sensing image Scene classification is an important research topic in remote sensing community and has evoked a growing concern with the recent development of deep learning techniques. However, the requirement of a large amount of annotations brings great challenges to deep learning-based scene classification approaches. Visual-linguistic pretraining models, which improve the transferability of visual models using the supervision information of text, create a new way for the task under label scarcity scenario. In this paper, we explore the novel approach of prompt engineering, aiming to achieve satisfactory performance of multi-modal pretraining models on downstream remote sensing image scene classification task with minimal amounts of training data. Experiments were conducted on multiple publicly available datasets. The results indicate that training the learnable prompts with a small number of samples can yield impressive results, surpassing the few-shot transfer learning results of the best-performing pre-trained models.