Spectral-spatial wave and frequency interactive transformer for hyperspectral image classification
Abstract Efficient extraction of spectral-spatial features is essential for accurate hyperspectral image (HSI) classification, where capturing both local texture and global semantic relationships is critical. While Convolutional Neural Networks (CNNs) and Transformers have shown strong capabilities...
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Nature Portfolio
2025-07-01
|
| Series: | Scientific Reports |
| Subjects: | |
| Online Access: | https://doi.org/10.1038/s41598-025-12489-3 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract Efficient extraction of spectral-spatial features is essential for accurate hyperspectral image (HSI) classification, where capturing both local texture and global semantic relationships is critical. While Convolutional Neural Networks (CNNs) and Transformers have shown strong capabilities in modeling local and global dependencies, most existing architectures operate directly on raw spectral-spatial inputs and lack explicit mechanisms for frequency-domain decomposition thereby overlooking potentially discriminative phase and frequency components. To address this limitation, we propose a Spectral-Spatial Wave and Frequency Interactive Transformer for HSI classification, which integrates frequency-aware and phase-aware token representations into a unified Transformer framework. Specifically, our model first employs a CNN backbone to extract shallow spectral-spatial features. These are then processed by a novel Frequency Domain Transformer Encoder, composed of two complementary branches: (i) a Spectral-Spatial Frequency Generator that extracts multiscale frequency features, and (ii) a Spectral-Spatial Wave Generator that encodes phase and amplitude characteristics as complex-valued wave tokens. A Spectral-Spatial Interaction Module fuses these components, followed by a Local-Global Modulator that refines semantic representations from multiple perspectives. Extensive experiments on five benchmark HSI datasets, demonstrate the effectiveness of our approach. The proposed model achieves state-of-the-art classification performance, with Overall Accuracies of 98.49%, 98.60%, 99.07%, 98.29%, and 97.97%, consistently outperforming existing methods. |
|---|---|
| ISSN: | 2045-2322 |