Self-Supervised Multiscale Contrastive and Attention-Guided Gradient Projection Network for Pansharpening

Area of Science:

• Remote Sensing
• Image Processing
• Computer Vision

Background:

• Pansharpening is vital for enhancing remote sensing image resolution.
• Deep learning methods are becoming dominant in image processing tasks.
• Existing pansharpening techniques face challenges in balancing spectral and spatial information.

Purpose of the Study:

• To develop an advanced deep learning model for pansharpening.
• To address the limitations of current methods in spectral-spatial information fusion.
• To improve the visual quality and quantitative metrics of pansharpened remote sensing images.

Main Methods:

• Formulating pansharpening as two optimization subproblems.
• Utilizing multiscale contrastive learning combined with attention-guided gradient projection networks.
• Designing a Spectral-Spatial Universal Module (SSUM) for spectral and spatial enhancement.
• Employing discrete wavelet transform (DWT) for feature extraction.
• Integrating contrastive learning and residual connections for information balancing.

Main Results:

• The proposed method demonstrates superior performance in visual quality.
• Quantitative evaluation metrics confirm the effectiveness of the approach.
• The method successfully generates high-resolution multispectral images.

Conclusions:

• The developed pansharpening technique offers significant improvements over existing methods.
• The integration of multiscale contrastive learning and attention mechanisms is effective.
• This research contributes a robust solution for high-quality remote sensing image fusion.