Few-shot Remote Sensing Scene Image Classification Method Based On Cross-Scale Efficient Hybrid Encoder

The field of few-shot remote sensing image classification often suffer from heavy data dependency and neglected inter / intra class relationships due to limited samples, leading to sub-optimal accuracy in existing methods. To address this critical issue, in this paper, we propose a novel framework based on a novel cross-scale efficient hybrid encoder and an adaptive complementary metric distance classifier, aiming to enhance the model's feature extraction capabilities and its ability to differentiate between inter-class and intra-class relationships. First, we propose a novel Cross-Scale Efficient Hybrid Encoder (EHE) transforms multi-scale features into image feature sequences through intra-scale interaction and cross-scale fusion, efficiently integrating fine-grained geometric details with high-level semantics into unified discriminative representations. .Secondly, a dynamic adaptive complementary metric distance classifier (ACMDC) is designed, which utilizing an improved adaptive cosine classifier for query matching and then updating the adaptive activation function, while the Euclidean distance and cosine similarity are used as complementary distance metrics, This adaptive fusion optimizes metric space by enforcing tighter intra-class clusters and separable inter-class boundaries.Experimental results on the NWPU-RESISC45, WHU-RS19, and UC Merced remote sensing datasets demonstrate that our method achieves accuracies of 72.45%, 83.69%, and 61.57% respectively under the 5-way 1-shot setting, outperforming the second-best model by 0.32%, 1.50%, and 1.39%. In the 5-way 5-shot configuration, our approach yields accuracy improvements of 2.57%, 0.86%, and 2.12% over the sub-optimal model. These findings confirm that the proposed method enables models to learn richer intra-class and inter-class relationships, significantly enhancing the discriminative capability of few-shot remote sensing scene classification models.