An Effective Hyperspectral Image Classification Network Based on Multi-Head Self-Attention and Spectral-Coordinate...

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Hyperspectral image (HSI) classification is crucial for analyzing spectral data.
• Convolutional Neural Networks (CNNs) show promise but struggle with accuracy and efficiency due to limited receptive fields and deep architectures.
• Existing methods often face challenges in balancing performance and computational load for HSI classification.

Purpose of the Study:

• To propose an effective hyperspectral image classification network that overcomes the limitations of CNN-based methods.
• To enhance both the accuracy and efficiency of HSI classification.
• To introduce a novel network architecture that integrates multi-head self-attention and spectral-coordinate attention.

Main Methods:

• A point-wise convolution network (PCN) is utilized to reduce spectral redundancy and improve discriminability.
• A modified multi-head self-attention (M-MHSA) model with down-sampling is employed to capture long-range dependencies efficiently.
• A lightweight spectral-coordinate attention fusion module combining spectral attention (SA) and coordinate attention (CA) is introduced to enhance feature weighting and object localization.

Main Results:

• The proposed MSSCA network demonstrates competitive performance on Indian Pines (IP), Pavia University (PU), and Salinas HSI datasets.
• Experimental results indicate significant improvements in classification accuracy compared to existing methods.
• The method achieves these accuracy gains without an increase in network complexity or computational cost.

Conclusions:

• The proposed multi-head self-attention and spectral-coordinate attention network (MSSCA) offers an effective solution for accurate and efficient HSI classification.
• The integration of PCN, M-MHSA, and spectral-coordinate attention fusion module successfully addresses the limitations of traditional CNNs.
• The method presents a highly competitive approach for HSI classification tasks, balancing performance and computational efficiency.