Leveraging potential of limpid attention transformer with dynamic tokenization for hyperspectral image classification

Area of Science:

• Remote Sensing
• Computer Vision
• Deep Learning

Background:

• Hyperspectral imaging (HSI) provides rich spectral information but faces challenges in spatial-spectral feature extraction.
• Convolutional Neural Networks (CNNs) excel at contextual modeling but struggle with global spatial-spectral correlations in HSI.
• Vision Transformers (ViTs) offer global context but can be computationally intensive and require effective positional encoding.

Purpose of the Study:

• To propose a novel deep learning architecture, the Limpid Size Attention Network (LSANet), for enhanced hyperspectral image analysis.
• To improve the extraction and representation of both spatial and spectral features in HSI.
• To develop a computationally efficient attention mechanism for capturing global spatial-spectral feature correlations.

Main Methods:

• LSANet integrates 3D and 2D convolution blocks to enhance spatial-spectral feature representation.
• A novel Limpid Attention Block (LAB) is introduced, utilizing LS attention for global spatial-spectral feature correlation.
• Conditional Position Encoding (CPE) is employed within the ViT encoder to dynamically generate tokens for richer contextual representation.

Main Results:

• LSANet achieved high overall accuracy (OA) on benchmark HSI datasets: 98.78% (IP), 98.67% (PU), 97.52% (SV), and 89.45% (Botswana).
• The proposed model demonstrated superior performance compared to existing CNN and transformer-based methods.
• LSANet effectively captures global spatial-spectral feature correlations with reduced computational cost compared to MHSA.

Conclusions:

• LSANet offers a significant advancement in hyperspectral image classification by effectively integrating spatial and spectral information.
• The proposed LS attention mechanism provides an efficient alternative for capturing global dependencies in HSI data.
• LSANet presents a promising deep learning approach for various remote sensing applications requiring high-accuracy HSI analysis.