GAFnet: Group Attention Fusion Network for PAN and MS Image High-Resolution Classification

Area of Science:

• Remote Sensing
• Computer Vision
• Image Processing

Background:

• Panchromatic (PAN) and multispectral (MS) images offer complementary spatial and spectral information for enhanced image interpretation.
• Existing methods may not fully leverage the synergistic potential of paired PAN and MS data.
• Accurate classification of fused image data is crucial for various applications.

Purpose of the Study:

• To propose a novel deep learning network for classifying fused PAN and MS images.
• To develop an effective feature extraction and fusion strategy that integrates spatial and spectral information.
• To enhance the accuracy and detail in image classification by leveraging complementary data sources.

Main Methods:

• A deep group spatial-spectral attention fusion network was developed for PAN and MS image classification.
• Multispectral images were unpooled to match the resolution of panchromatic images.
• Group spatial attention and group spectral attention modules were employed for feature extraction.
• An attention fusion module integrated multi-level features, preserving both global and local information.
• Pixel-level classification was performed using the fused features.

Main Results:

• The proposed deep group spatial-spectral attention fusion network demonstrated comparable results across four diverse datasets.
• The method effectively fused low-level and high-level features, maintaining both abstract and detailed information.
• Attention mechanisms successfully integrated spatial and spectral features for improved classification outcomes.
• Experimental validation confirmed the efficacy of the proposed approach.

Conclusions:

• The developed deep group spatial-spectral attention fusion network offers a robust solution for PAN and MS image classification.
• The attention-based fusion strategy effectively leverages complementary spatial and spectral information.
• The method shows significant potential for advancing remote sensing image analysis and interpretation.