Unsupervised Deep Tensor Network for Hyperspectral-Multispectral Image Fusion.

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Hyperspectral image (HSI) resolution enhancement via multispectral image (MSI) fusion is crucial.
• Current deep learning (DL) fusion methods face challenges in representing multidimensional HSI data and often require unavailable ground truth data.

Purpose of the Study:

• To propose an unsupervised deep tensor network (UDTN) for HSI-MSI fusion.
• To address the limitations of existing DL fusion techniques, particularly the need for ground truth data and effective multidimensional feature representation.

Main Methods:

• Integration of tensor theory with DL to develop a UDTN.
• Introduction of a coupled tensor filtering module for joint representation of LR HSI and HR MSI.
• Utilizing a co-attention mechanism within a projection module for unsupervised, end-to-end training.

Main Results:

• The UDTN effectively fuses LR HSI and HR MSI by representing spectral and spatial modes and their interactions.
• The proposed method achieves accurate HSI-MSI fusion without relying on HR HSI ground truth.
• Experimental validation on simulated and real remote-sensing datasets confirms the method's effectiveness.

Conclusions:

• The developed UDTN offers a robust solution for HSI-MSI fusion, overcoming key limitations of existing approaches.
• The unsupervised, end-to-end training framework enhances practical applicability in real-world scenarios.
• This tensor-based deep learning approach significantly advances HSI resolution enhancement technology.