.png){kind=icon}
.png){kind=icon}
Return
Relation-Aware Multiprototype Learning for Semi-Supervised Hyperspectral Image Classification
.png){kind=icon}
PRE
AI
DOI：10.1109/TGRS.2025.3626590.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Pseudo-label learning-based methods have achieved excellent performance by leveraging spatial–spectral features of massive unlabeled samples in hyperspectral image (HSI) classification. Existing methods primarily rely on confidence scores predicted by classification models to produce pseudo-labels. However, both limited labeled samples and complex spatial–spectral information result in these models being prone to overfitting, reducing the reliability of pseudo-labels. In this article, a relation-aware multiprototype learning (RAMPL) method is proposed. First, to extract discriminative spatial–spectral features, a hybrid transformer–CNN module (TCM) is developed, which extracts both global and local features using the attention mechanism and a mixed-scale convolutional block (MSCB). Then, to generate high-confidence pseudo-labels, a multiprototype learning module (MPLM) is proposed to construct multiple class prototypes for each category by exploiting both unlabeled samples with pseudo-labels and a small number of labeled samples, which effectively models the intraclass diversity. Finally, to further enhance feature discrimination, a relation-aware contrastive learning module (RACLM) is proposed to learn the complex spatial distribution difference and spectral variability of the HSI by constructing a relation-aware pseudo-graph. Extensive experiments demonstrate that the proposed RAMPL achieves superior classification accuracy, achieving 85.76%, 84.36%, 93.14%, and 96.13% on the Indian Pines (IP), Houston2013 (H2013), Salinas (SA), and Kennedy Space Center (KSC) datasets with five labeled samples in each category, and surpasses 11 state-of-the-art HSI classification methods.
Keywords:
Class prototype
hyperspectral image (HSI) classification
pseudo-label
semi-supervised learning
transformer
Journal
IF:
8.6
Papers:
2.1W
Citations:
10.7W
View more organizations