A Lightweight 1-D Convolution Augmented Transformer with Metric Learning for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Deep Learning

Background:

• Hyperspectral image (HSI) classification is crucial in remote sensing.
• Deep learning, particularly Convolutional Neural Networks (CNNs), has shown promise but demands significant computational resources.
• Transformers, successful in natural language processing, offer a potential alternative for image recognition tasks.

Purpose of the Study:

• To propose a novel deep learning model for HSI classification using transformers.
• To address the limitations of existing CNN-based methods regarding GPU memory and runtime.
• To enhance classification performance, especially in scenarios with limited training data.

Main Methods:

• A transformer-based model architecture is proposed for HSI classification.
• Metric learning is combined with the transformer model for the first time in this context.
• 1-D convolution and Mish activation function are incorporated to improve performance with limited samples.

Main Results:

• The proposed transformer-based model demonstrates superior accuracy compared to existing methods.
• The model shows reduced GPU memory requirements and faster running times.
• Experimental results on three HSI datasets validate the model's effectiveness.

Conclusions:

• The novel transformer-based approach offers significant advantages for HSI classification.
• This method provides a more efficient and accurate solution, particularly beneficial for limited data scenarios.
• The integration of metric learning and transformers opens new avenues for HSI analysis.