Few-shot Remote Sensing Scene Classification via Parameter-free Attention and Region Matching

Abstract

Few-shot remote sensing scene classification, a pivotal task in geospatial scene understanding, has drawn considerable attention as a means to address annotation scarcity in Earth observation. While recent advancements exploit metric-based learning, conventional methods that rely on global feature aggregation, e.g., prototype networks, often entangle target objects with cluttered backgrounds—an inherent limitation given the heterogeneous land-cover elements in remote sensing imagery. Although parametric attention mechanisms partially alleviate this issue, they tend to overfit base-class patterns, limiting adaptability to novel categories with diverse intra-class variations. To tackle these challenges, we propose the Parameter-free Attention with Selective Region Matching (PA-SRM) framework, which integrates two cascaded components: a parameter-free region attention module and a local description classifier. The former dynamically emphasizes discriminative regions by jointly assessing semantic similarity and spatial coherence. At the same time, the latter explicitly employs entropy-aware multi-region voting to suppress residual background interference in queries. Extensive experiments on NWPU-RESISC45, WHU-RS19, UCM, and AID datasets validate the superiority of PRA-SRM and the effectiveness of its components.