.png){kind=icon}
.png){kind=icon}
Return
Global Semantic Prototype-Guided Contrastive Learning for Few-Shot Joint Classification of MS and PAN
.png){kind=icon}
PRE
AI
朱
W
X
L
W
B
DOI：10.1109/tgrs.2026.3716678.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Self-supervised cross-source contrastive learning (CL) for few-shot joint classification of panchromatic (PAN) and multispectral (MS) images focuses on learning transferable representations from large-scale unlabeled multisource images. However, existing methods neglect semantic relationships among samples, leading to inaccurate definitions of positive–negative pairs. Moreover, remote sensing noise images may reduce the reliability of pseudo-label generation, which further affects cross-source alignment. In this paper, we propose a global semantic prototype-guided contrastive learning (GSPCL) method for few-shot joint classification of MS and PAN. In particular, we design a pseudolabel assignment strategy that uses globally optimized semantic prototypes to assign pseudolabels to unlabeled images. By defining positive and negative pairs based on semantic pseudolabels, this strategy enables more reliable cross-source alignment, effectively mitigates false negatives, and reduces the influence of local noise. To reduce the interference of unreliable pseudolabels, we introduce a dynamic selection strategy that selects reliable semantic pseudolabels based on confidence and stability metrics and uses them to guide cross-source alignment. Experimental results on the Hohhot, Nanjing, and Xi’an datasets show that GSPCL consistently outperforms existing state-of-the-art methods, improving the Kappa by 0.85%, 1.69%, and 2.70%, respectively. These results verify the effectiveness and robustness of GSPCL with limited labeled data. Our code is available at: <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/wanling0926/GSPCL</uri>
Keywords:
Contrastive learning (CL)
deep learning
few-shot
multispectral (MS) and panchromatic (PAN) images
semantic prototype
Journal
IF:
8.6
Papers:
2.1W
Citations:
10.7W
Organization
Cited Papers