GroupFormer for hyperspectral image classification through group attention.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image (HSI) data offers rich spectral information but faces challenges like limited training samples and redundant data.
• Convolutional Neural Networks (CNNs) extract low-level features but struggle with long-range dependencies in HSI data.
• Vision transformers effectively capture long-range dependencies using attention mechanisms, yet typically require substantial labeled data.

Purpose of the Study:

• To address the data scarcity issue in hyperspectral image classification.
• To develop a model capable of extracting both low-level and high-level features efficiently.
• To improve classification accuracy using limited training samples.

Main Methods:

• Proposed a spectral-spatial feature extractor group attention transformer architecture.
• Incorporated a multiscale feature extractor for shallow feature extraction.
• Introduced a group attention mechanism for high-level semantic feature extraction.

Main Results:

• The proposed model achieved state-of-the-art classification results on four public HSI datasets (Indian Pines, Pavia University, Salinas, KSC).
• Achieved superior performance in Overall Accuracy (OA), Average Accuracy (AA), and Kappa coefficient.
• Demonstrated effectiveness using minimal training samples (5%, 1%, 1%, and 10% across datasets).

Conclusions:

• The spectral-spatial feature extractor group attention transformer effectively handles HSI data challenges, particularly data scarcity.
• The model's hybrid approach of multiscale extraction and group attention enhances feature representation.
• This method offers a promising solution for accurate HSI classification with limited labeled data.