A 3D-2D Multibranch Feature Fusion and Dense Attention Network for Hyperspectral Image Classification

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image classification (HSI) is crucial for analyzing Earth's surface.
• Convolutional neural networks (CNNs) show promise but suffer from underutilized spectral-spatial features and redundant information.
• Existing CNNs face challenges with convergence difficulty and feature extraction limitations.

Purpose of the Study:

• To propose a novel 3D-2D multibranch feature fusion and dense attention network for HSI classification.
• To overcome the limitations of existing methods, including underutilization of spectral-spatial features and convergence issues.
• To enhance the accuracy and robustness of hyperspectral image classification.

Main Methods:

• A 3D multibranch feature fusion module integrates multiple receptive fields for shallow feature extraction.
• A 2D densely connected attention module employs densely connected layers to prevent gradient vanishing and improve feature reuse.
• A spatial-channel attention block within the 2D module emphasizes key features and suppresses noise across spatial and channel dimensions.

Main Results:

• The proposed network demonstrates significant improvements in hyperspectral image classification performance.
• Experimental results on four benchmark datasets confirm the model's effectiveness and robustness.
• The method successfully addresses issues of underutilized spectral-spatial features and redundant information.

Conclusions:

• The novel 3D-2D multibranch dense attention network offers a robust solution for HSI classification.
• The proposed architecture effectively integrates spectral and spatial information for superior feature representation.
• This approach enhances classification accuracy and overcomes common challenges in hyperspectral imaging analysis.