SF2SS: a few-shot hyperspectral image classification method based on spectral band attention and two-stage prototype separation

Hyperspectral Image (HSI) classification faces significant challenges in data-scarce scenarios, particularly in Few-Shot Learning (FSL), where issues like class imbalance and spectral band redundancy further increase classification difficulty. This paper proposes a novel few-shot hyperspectral image classification method called SF2SS (Spectral Focus 2 Stage Separation), which integrates spectral band attention and a two-stage prototype separation strategy. The method employs a class-specific spectral attention mechanism to dynamically identify the spectral bands that are most decisive for distinguishing each class. It also utilizes a global spectral attention mechanism to amplify the impact of bands that are highly discriminative across classes. Furthermore, a two-stage prototype separation strategy is introduced: first, prototypes of confusing large-sample classes are separated using a confusion matrix; then, prototypes of small-sample classes are further refined, aiming to alleviate the instability of prototypes under few-shot conditions. Experiments conducted on the Indian Pines (IP) and Salinas (SA) datasets demonstrate that SF2SS achieves excellent performance in few-shot scenarios. It reaches over 97% overall accuracy (OA) on the IP dataset and over 98% on the SA dataset in the 1-shot setting, significantly outperforming traditional machine learning methods (e.g., SVM and Random Forest) as well as few-shot learning methods (e.g., Prototypical Networks and DCFSL). SF2SS particularly excels in large-sample and medium sample classes, while also achieving notable improvements in small-sample classes. These results validate the effectiveness of the proposed spectral attention mechanisms and the two-stage prototype separation strategy.