CNN-Transformer and Channel-Spatial Attention based network for hyperspectral image classification with few samples.

Area of Science:

• Earth observation
• Remote sensing
• Computer vision

Background:

• Deep learning models for hyperspectral image classification typically require extensive annotated data.
• Acquiring sufficient annotated hyperspectral imagery from remote or high-altitude regions presents significant challenges.
• Limited labeled data hinders the development and application of advanced classification techniques.

Purpose of the Study:

• To propose a novel algorithm, CTA-net, designed for effective hyperspectral image classification with minimal annotated samples.
• To address the data scarcity issue in hyperspectral image analysis for Earth observation.
• To enhance the performance of deep learning models in challenging remote sensing scenarios.

Main Methods:

• A sample expansion scheme was developed to generate synthetic data, mitigating the insufficient sample problem.
• A novel deep learning architecture, CTA-net, was introduced, integrating Convolutional Neural Networks (CNN) and Transformer modules.
• The network employs CNN for local feature extraction and Transformer for non-local feature analysis, further optimized by a channel-spatial attention module.

Main Results:

• Experiments conducted on multiple hyperspectral image datasets demonstrated the efficacy of the proposed CTA-net algorithm.
• The sample expansion technique successfully alleviated the limitations posed by scarce annotated data.
• The combined CNN-Transformer approach effectively captured both local and non-local features for improved classification.

Conclusions:

• CTA-net offers a viable solution for hyperspectral image classification in data-scarce environments.
• The integration of CNN and Transformer, coupled with sample expansion, significantly enhances classification performance.
• This research contributes to advancing Earth observation and remote sensing capabilities through efficient deep learning.