DMAF-NET: Deep Multi-Scale Attention Fusion Network for Hyperspectral Image Classification with Limited Samples

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Deep learning, particularly Convolutional Neural Networks (CNNs), has shown success in hyperspectral image classification (HSIC).
• A major challenge in HSIC is the scarcity of labeled training data, which limits the accuracy and generalization of CNN models.
• Existing methods struggle to effectively leverage deep features from limited samples.

Purpose of the Study:

• To propose a novel deep multi-scale attention fusion network (DMAF-NET) for enhanced HSIC.
• To address the challenge of limited labeled samples in HSIC tasks.
• To improve classification accuracy and generalization ability using multi-scale feature extraction and attention mechanisms.

Main Methods:

• Designed a baseline network with a pyramid structure and densely connected 3D octave convolutions for multi-scale feature extraction.
• Developed a multi-scale spatial-spectral attention module and a pyramidal multi-scale channel attention module to capture complex dependencies.
• Integrated a multi-attention fusion module to effectively combine features from different branches.

Main Results:

• The proposed DMAF-NET achieved high classification accuracy on four benchmark datasets.
• The method demonstrated effectiveness even when trained with a limited number of labeled samples.
• The attention fusion strategy successfully integrated multi-scale and multi-level features.

Conclusions:

• The DMAF-NET effectively enhances hyperspectral image classification performance, especially under limited labeled data conditions.
• The integration of multi-scale features and attention mechanisms is crucial for improving HSIC accuracy.
• The proposed network offers a promising solution for practical HSIC applications with data constraints.