TUH-NAS: A Triple-Unit NAS Network for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Neural Architecture Search (NAS) has shown promise in hyperspectral image classification.
• Existing NAS methods often overlook the intricate relationship between spatial and spectral data in hyperspectral imagery.
• Effective integration of spatial and spectral features is vital for enhancing classification accuracy.

Purpose of the Study:

• To introduce a novel Triple-Unit Hyperspectral NAS (TUH-NAS) network for improved hyperspectral image classification.
• To enhance the fusion of spatial and spectral information within the classification network.
• To boost the overall classification efficacy, particularly in recognizing object boundaries.

Main Methods:

• Development of a Triple-Unit Hyperspectral NAS (TUH-NAS) network.
• Implementation of a fusion unit to strengthen the intrinsic relationship between spatial and spectral information.
• Design of a hyperspectral image attention mechanism module.
• Adoption of a composite loss function to focus on difficult-to-classify samples.

Main Results:

• The TUH-NAS network demonstrated superior performance in hyperspectral image classification compared to existing NAS methods.
• The proposed method showed improved recognition of object boundaries, even with limited training samples.
• Experimental evaluations on three public hyperspectral datasets validated the effectiveness of TUH-NAS.

Conclusions:

• The TUH-NAS network effectively integrates spatial and spectral features for enhanced hyperspectral image classification.
• The attention mechanism and composite loss function contribute to improved accuracy and focus on challenging regions.
• TUH-NAS represents a significant advancement in NAS-based hyperspectral image classification, particularly for boundary recognition.