.png){kind=icon}
.png){kind=icon}
Return
A Unified Self-Supervised Learning Framework for Hyperspectral Image Classification
.png){kind=icon}
OA
AI
DOI：10.1109/ACCESS.2025.3549277.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Self-supervised learning (SSL) methods, including contrastive learning (CL) and masked image modeling (MIM), have shown commendable performance on various remote sensing visual tasks. However, learning comprehensive and complementary representations from unlabeled remote sensing images remains a challenging task due to their intricate spatial-spectral characteristics. To address this issue, we propose a Unified Self-supervised Learning Framework (USeLF) that incorporates both CL and MIM to enhance hyperspectral image (HSI) classification. USeLF employs an asymmetric Siamese network framework, designed to accommodate both CL and MIM. Leveraging the common characteristics of these two paradigms, we developed a hierarchical Transformer structure that integrates downsampling and upsampling modules to improve model generalization. Furthermore, image masking is devised as a novel data augmentation strategy to bridge the gap between CL and MIM. USeLF learns diverse and complementary features from both CL and MIM, resulting in superior performance for HSI classification. Experiments conducted on four widely used HSI datasets demonstrate that USeLF outperforms existing state-of-the-art (SOTA) methods. Notably, on the WHU-Hi-HongHu dataset-a complex agricultural scene with numerous crop classes-USeLF achieves a 5.3% improvement in overall accuracy compared to the current SOTA method.
Keywords:
contrastive learning
masked image modeling
hyperspectral image classification
Self-supervised learning
Self-supervised learning
contrastive learning
masked image modeling
hyperspectral image classification
Journal
IF:
3.6
Papers:
9.8W
Citations:
29.4W