A Multi-Stage Progressive Pansharpening Network Based on Detail Injection with Redundancy Reduction

Area of Science:

• Remote Sensing
• Image Processing
• Computer Vision

Background:

• Pansharpening enhances multispectral images using panchromatic data.
• Deep learning offers potential for improving pansharpening quality.

Purpose of the Study:

• To propose a novel deep learning-based pansharpening method.
• To enhance both spatial and spectral details in fused images.

Main Methods:

• Introduced the Multi-Stage Progressive Pansharpening Network with Detail Injection with Redundancy Reduction Mechanism (MSPPN-DIRRM).
• Employed a three-level network for multi-scale spectral and spatial feature extraction.
• Introduced a Redundancy Reduction Mechanism (DRRM) for image reconstruction.

Main Results:

• The MSPPN-DIRRM model demonstrated superior performance over existing deep learning methods.
• Experimental results on simulated and real satellite data (QuickBird, GaoFen1, WorldView2) validated the model's effectiveness.
• Achieved performance improvements of 0.92-18.7% across various evaluation metrics.

Conclusions:

• The proposed MSPPN-DIRRM method effectively improves pansharpening quality.
• The DRRM module successfully reduces spatial and channel redundancy, enhancing fusion.
• The method shows significant advancements in remote sensing image enhancement.