Patch-based Cross-domain Adaptation for Few-shot Remote Sensing Scene Recognition

Recognizing new classes from a few labeled remote sensing images presents a significant challenge. Existing methodologies assume that training and testing datasets come from identical domains. Collecting training datasets of remote sensing images from each domain is practically unfeasible. To tackle this challenge, we propose a strategy that leverages image patch feature embedding and cross-adaptation mechanism for cross-domain few-shot classification. Initially, the supplied image is segmented into patches of the same size. Subsequently, feature embeddings are extracted and reinforced with prototypical triplet loss to enhance robustness. A unified module then facilitates cross-adaptation between the support and query sets, producing classification features that effectively represent the relationship between them, thus better adapting to few-shot remote sensing classification challenges with domain shifts. Our proposed method’s excellence and generality are confirmed by comprehensive experiments.