Uncertainty-driven mixture convolution and transformer network for remote sensing image super-resolution.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) and Transformer-based Networks show promise in remote sensing image super-resolution (RSISR).
• Effective fusion of CNN inductive bias and Transformer long-range modeling for RSISR is underexplored.

Purpose of the Study:

• To propose an uncertainty-driven mixture convolution and transformer network (UMCTN) for improved RSISR performance.
• To enhance reconstruction quality, particularly in texture and edge regions.

Main Methods:

• UMCTN employs a U-shape architecture for multi-scale and hierarchical feature acquisition.
• A novel dense-sparse transformer group (DSTG) is introduced in the latent layer to mitigate quadratic complexity.
• An uncertainty-driven loss (UDL) focuses network attention on high-variance pixels.

Main Results:

• UMCTN achieves state-of-the-art performance on UCMerced LandUse and AID datasets.
• The proposed network demonstrates superior reconstruction quality in challenging texture and edge areas.
• The DSTG effectively addresses the computational complexity of standard Transformer models.

Conclusions:

• UMCTN offers a significant advancement in remote sensing image super-resolution.
• The integration of CNNs and Transformers, guided by uncertainty, yields superior results.
• The method shows strong potential for practical applications in remote sensing image enhancement.