Spectral-spatial wave and frequency interactive transformer for hyperspectral image classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Accurate hyperspectral image (HSI) classification requires efficient spectral-spatial feature extraction.
• Existing methods often overlook discriminative frequency-domain components by operating on raw inputs.
• Convolutional Neural Networks (CNNs) and Transformers excel at local and global dependencies, respectively, but lack explicit frequency analysis.

Purpose of the Study:

• To develop a novel Spectral-Spatial Wave and Frequency Interactive Transformer for HSI classification.
• To integrate frequency-aware and phase-aware token representations into a unified Transformer framework.
• To overcome the limitations of existing architectures by incorporating explicit frequency-domain decomposition.

Main Methods:

• Utilized a CNN backbone for initial spectral-spatial feature extraction.
• Developed a Frequency Domain Transformer Encoder with complementary Spectral-Spatial Frequency and Wave Generators.
• Employed a Spectral-Spatial Interaction Module and Local-Global Modulator for feature fusion and refinement.

Main Results:

• The proposed model achieved state-of-the-art classification performance on five benchmark HSI datasets.
• Demonstrated high Overall Accuracies: 98.49%, 98.60%, 99.07%, 98.29%, and 97.97%.
• Consistently outperformed existing HSI classification methods.

Conclusions:

• The integration of frequency and phase information significantly enhances spectral-spatial feature representation.
• The proposed Spectral-Spatial Wave and Frequency Interactive Transformer offers a powerful new approach for HSI classification.
• The model's effectiveness is validated by its superior performance across multiple datasets.