The article presents a novel method for cross-domain few-shot hyperspectral image classification. The proposed method differs from current ones by learning the relations between samples from different views and incorporating them into the model learning process, rather than simply predicting classes based on distance to support samples or prototypes. The method is built on the DCFSL method, applying contrastive learning to learn class-level sample relations for more discriminable sample features and a transformer-based cross-attention learning module to learn set-level sample relations. Experimental results demonstrate the effectiveness of the multi-view relation learning mechanism compared with existing methods.
Publication date: 2 Nov 2023
Project Page: https://github.com/HENULWY/STBDIP
Paper: https://arxiv.org/pdf/2311.01212
