Dynamic gated fusion network with hierarchical multi-scale attention for hyperspectral image classification

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image classification faces challenges with fixed feature fusion strategies.
• Existing methods struggle with adapting to diverse data characteristics and integrating multi-scale features with attention mechanisms effectively.

Purpose of the Study:

• To propose a novel dynamic gated fusion network with hierarchical multi-scale attention (DGFNet) for improved hyperspectral image classification.
• To address the limitations of fixed fusion strategies and enhance the synergy between multi-scale feature extraction and attention mechanisms.

Main Methods:

• The DGFNet employs a multi-scale feature aggregator (MSFA) using pyramid expansion convolution for comprehensive spatial feature extraction.
• An enhanced channel-spatial attention (ECSA) module with multi-pooling and cascaded structures deepens feature interaction.
• A dynamic gated fusion module adaptively adjusts feature contributions based on data characteristics.

Main Results:

• DGFNet achieved high accuracy rates on benchmark datasets: Pavia University (96.91%), Houston (97.12%), Indian Pines (94.05%), and WHU-HongHu (94.46%).
• Dynamic gated fusion outperformed other fusion strategies in classification accuracy, computational efficiency, and model stability.
• Ablation experiments confirmed the effectiveness and necessity of each proposed module.

Conclusions:

• DGFNet offers an efficient, accurate, and robust solution for hyperspectral image classification.
• The dynamic gated fusion approach provides superior performance compared to traditional fusion methods.
• The proposed method demonstrates significant advancements in adapting to various remote sensing data characteristics.