FrMLNet: Framelet-Based Multilevel Network for Pansharpening.

Area of Science:

• Remote Sensing
• Computer Vision
• Image Processing

Background:

• Satellites provide panchromatic (PAN) and multispectral (MS) images with complementary spatial and spectral resolutions.
• Pansharpening aims to fuse PAN and MS images for high-resolution multispectral data, but often struggles with simultaneous spatial and spectral preservation.
• Existing methods face challenges in effectively integrating high spatial detail from PAN images and rich spectral information from MS images.

Purpose of the Study:

• To develop an advanced pansharpening method capable of achieving both high spatial and high spectral resolution.
• To propose a novel framelet-based convolutional neural network (CNN) architecture for improved pansharpening performance.
• To overcome the limitations of conventional pansharpening techniques in preserving both spatial and spectral information.

Main Methods:

• A framelet-based CNN architecture comprising three subnetworks: feature embedding, feature fusion, and framelet prediction.
• Predicting framelet coefficients of high-resolution MS images rather than directly inferring the images.
• Incorporating multilevel feature aggregation and hybrid residual connections to leverage PAN and MS image information.

Main Results:

• The proposed framelet-based CNN method demonstrates superior performance in quantitative and qualitative evaluations.
• Achieved more appealing results compared to existing state-of-the-art pansharpening methods.
• Successfully preserved both spatial and spectral information simultaneously in the fused images.

Conclusions:

• The framelet-based CNN approach offers a significant advancement in pansharpening technology.
• The method effectively enhances both spatial and spectral resolution in satellite imagery.
• The developed technique provides a robust solution for obtaining high-quality fused multispectral images.