A dual-channel hyperspectral classification method based on NAS and transformer.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Transformer networks excel at capturing long-range dependencies globally.
• Neural Architecture Search (NAS) has advanced hyperspectral image (HSI) classification by focusing on local information.
• Existing two-channel networks struggle to simultaneously leverage both local and global features in HSI data, limiting classification accuracy.

Purpose of the Study:

• To propose CTmixer-NAS, a novel two-channel network for HSI classification.
• To simultaneously capture local and global information for improved HSI classification accuracy.
• To leverage NAS for automated design of high-performance CNN components within the network.

Main Methods:

• Developed CTmixer-NAS, a two-channel network integrating Transformer and CNN branches.
• Implemented information fusion across self-coding layers in the Transformer branch.
• Employed NAS to automatically design an optimal CNN structure for the CNN branch.

Main Results:

• CTmixer-NAS achieved superior performance across five benchmark hyperspectral datasets.
• The proposed method effectively integrates local and global information, overcoming limitations of previous approaches.
• NAS-driven CNN design enhanced classification accuracy and generalization performance.

Conclusions:

• CTmixer-NAS offers a robust solution for HSI classification by synergistically combining Transformer and NAS-optimized CNNs.
• The approach demonstrates significant improvements in accuracy and generalization, reducing manual labor.
• This method sets a new standard for HSI classification by effectively processing both local and global spectral-spatial information.