Abstract:

In response to the challenges in the field of cross-domain few-shot classification of hyperspectral images,such as low classification accuracy and limited generalization capability,this study proposes a novel hyperspectral image classification method based on the subspace and memory bank of cross-domain few-shot learning(SMB-CFSL).A feature extractor is improved that integrates the channel attention mechanism and the spectral-spatial attention mechanism to fully extract the spectral spatial information on original hyperspectral images.By employing the contrastive learning mechanism to analyze the diversity and differences among small samples,the discriminative power and generalization performance of the model are enhanced under the few-shot scenario.Additionally,the prototype network is improved by utilizing adaptive subspace to enhance the utilization of embedding features,leading to improved accuracy in image classification.Finally,a memory bank module is introduced to achieve cross-domain alignment and enhance the classification performance of the model under cross-domain conditions.Through iterative training and continuous optimization,the optimized feature extractor is employed for classification on the testing set.We compare our proposed method with state-of-the-art approaches for cross-domain few-shot classification of hyperspectral images using four widely adopted datasets.Experimental results demonstrate that our method outperforms several existing methods in classification while also exhibiting excellent generalization capability and robustness.

Key words: image classification, cross-domain few-shot, feature extraction, subspace, memory bank