Hyperspectral image classification by integrating attention-based LSTM and hybrid spectral networks

ABSTRACT

Though hyperspectral remote sensing images contain rich spatial and spectral information, they pose challenges in terms of feature extraction and mining. This paper describes the integration of a dimensionality reduction technique that employs spectral attention and Hybrid Spectral Networks (HybridSN) with spatial attention for hyperspectral image classification. The goal of this approach is to improve the ability to classify hyperspectral images by increasing the capabilities of spectral-spatial feature fusion. Experiments on three hyperspectral datasets (Indian Pines, University of Pavia, and Houston University) demonstrate that our method’s overall accuracy is 99.66%, 99.97%, and 99.17% under 20% of the training samples, respectively, which is superior to several well-known approaches.

KEYWORDS:

• Convolutional neural network
• hyperspectral image classification
• LSTM autoencoder
• spectral-spatial attention

Acknowledgments

The authors would also like to extend a special acknowledgment to the anonymous reviewers for their expertise in improving this research. The researchers also would like to acknowledge the Universiti Putra Malaysia for the resources support and the Iraqi Ministry of Environment, Al-Anbar Environment Directorate for the kind financial support for this research.

Disclosure statement

No potential conflict of interest was reported by the author(s).