Few-Shot Remote Sensing Scene Recognition via Feature Enhancing Learning

Inspired by the efficient knowledge assimilation of the human brain, few-shot learning presents significant promise in machine learning and artificial intelligence (AI). This field aims to create models that classify and recognize new categories using a few labeled samples. However, the capacity of existing models to utilize knowledge from benchmark datasets requires further enhancement. Our research introduces an innovative model for few-shot scene recognition in optical remote-sensing imagery. This model incorporates a cross-attention mechanism to improve feature extraction with limited data. Additionally, it employs triplet angular loss to introduce geometric constraints in the feature space. Our method also combines non-linear projection transformations with fine-grained, instance-level classification. This enhances separation between classes and improves prediction accuracy. We have evaluated our model using three renowned datasets for few-shot remote sensing imagery: Northwestern Polytechnical University (NWPU), the Aerial Image Dataset (AID), and the University of California, Merced (UCM). The outcomes affirm our method’s capability to feature across diverse remote sensing scene categories. Our approach represents advancing precision and efficacy of detecting category-specific features in few-shot learning contexts.