Cross-scale feature representation and adaptive complementary metric learning for few-shot remote sensing scene classification

Abstract Few-shot remote sensing classification remains a challenging task due to the severe data scarcity and the insufficient modeling of intra-class compactness and inter-class separability in existing methods, Which often resulting in sub-optimal classification performance. To address this critical issue, in this paper, we propose a novel few-shot classification framework, namely CAME-Net, which incorporating a Cross-Scale Efficient Hybrid Encoder (EHE) and an Adaptive Complementary Metric Distance Classifier (DACMDC) to learn more discriminative and structurally consistent feature representations. Firstly, we propose a novel Cross-Scale Efficient Hybrid Encoder (EHE), that effectively captures cross-scale contextual information by transforming multi-scale features into image-level feature sequences through intra-scale interaction and cross-scale fusion, thereby integrating fine-grained geometric details with high-level semantics into unified discriminative representations. Secondly, a dynamic adaptive complementary metric distance classifier (DACMDC) is designed, which to adaptively optimize the metric space by jointly exploiting Euclidean distance and cosine similarity as complementary metrics, governed by a dynamic activation update mechanism. This dual-metric approach promotes tighter intra-class clustering and more distinct inter-class boundaries. Extensive experiments on three benchmark datasets—NWPU-RESISC45, WHU-RS19, and UC Merced—demonstrate the superiority of the proposed method. Under the 5-way 1-shot setting, our approach achieves accuracies of 72.45%, 83.69%, and 61.57%, outperforming SOTA methods by 0.32%, 1.50%, and 1.39%, respectively. In the 5-way 5-shot scenario, accuracy improvements reach up to 2.57%. These results confirm that the proposed framework significantly enhances feature representation and similarity measurement for complex remote sensing scenes.