Deep Multiscale Detail Networks for Multiband Spectral Image Sharpening.

Area of Science:

• Remote Sensing
• Computer Vision
• Deep Learning

Background:

• Pansharpening aims to enhance spatial resolution of multispectral images using high-resolution panchromatic images.
• Existing methods often struggle with preserving both spectral fidelity and spatial detail.

Purpose of the Study:

• Introduce a novel deep detail network for advanced image sharpening.
• Improve pansharpening performance by preserving spectral and spatial information.

Main Methods:

• Developed an end-to-end network fusing low-resolution multispectral and panchromatic inputs.
• Utilized grouped multiscale dilated convolutions for efficient multiscale feature extraction.
• Trained the network in the high-frequency domain for enhanced spatial preservation.
• Removed pooling and batch normalization layers to maintain spatial information and improve generalization.

Main Results:

• Achieved superior performance in both qualitative and quantitative assessments compared to existing methods.
• Demonstrated effective generalization across different satellite datasets without parameter tuning.
• Successfully reconstructed residual images containing fine spatial details from panchromatic data.

Conclusions:

• The proposed network offers a robust and generalizable solution for multiband spectral image sharpening.
• The architecture effectively balances spectral and spatial preservation for high-quality remote sensing imagery.
• The model framework is adaptable for other applications like hyperspectral image sharpening.