Dynamic Task-Oriented Prompting for Generalized Aerial Category Discovery

Remote sensing plays a crucial role in monitoring Earth’s surface and atmosphere, supplying valuable data on environmental factors such as land cover and vegetation health, which are essential for decision-making in agriculture, urban planning, disaster management, and other fields. With the advent of AI, researchers are striving to automate remote sensing image analysis by leveraging large pre-trained foundation models. However, adapting these models to remote sensing imagery poses significant challenges, particularly in fine-tuning them with sufficient labeled data, which can be costly and impractical. Moreover, existing methods may struggle to identify unknown categories in rapidly evolving environments. To tackle these issues, we have developed Dynamic Task-oriented Prompting (DTP), a technique that enables foundation models to effectively recognize both known and unknown scene categories in remote sensing imagery using limited labeled data. Extensive experiments have shown its promise in adapting foundation models to remote sensing imagery within dynamic environments.