Fusing 3D-CNN and lightweight Swin Transformer networks for HSI

Baisen Liu, Yuanjia Liu, Wulin Zhang, Yiran Tian

Keywords: Hyperspectral Image classification, Swin transformer, CNN, Deep Learning

TL;DR: Hyperspectral image classification using a combination of CNN feature extraction and swin transformer attention mechanism

Abstract: Recently deep learning has occupied an important position in hyperspectral image (HSI) classification. In this study, we explore the advantages of using convolutional neural networks (CNN) for feature extraction and fusing an advanced shift-window (swin) transformer network based on the transformer model for HSI classification. The swin transformer network attention perception, capable of learning local and global features, can avoid the dependence on single features during HSI classification. The experiments show that our proposed model outperforms traditional machine learning models, and achieves competitive results with advanced models.