EDB-Net: Efficient Dual-Branch Convolutional Transformer Network for Hyperspectral Image Classification
Hyperspectral image (HSI) classification, as a pivotal technology in remote sensing data processing, has garnered significant attention in recent years. Deep learning (DL) has been widely adopted for HSI classification due to its superior feature extraction capabilities. Nevertheless, the deployment...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10989234/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Hyperspectral image (HSI) classification, as a pivotal technology in remote sensing data processing, has garnered significant attention in recent years. Deep learning (DL) has been widely adopted for HSI classification due to its superior feature extraction capabilities. Nevertheless, the deployment of most existing DL models on resource-constrained devices remains challenging because of their intricate architectures and high computational demands. To tackle this challenge, we propose a lightweight dual-branch convolutional transformer network with efficient attention-aware mechanism (EDB-Net), which aims to balance model complexity, classification accuracy, and inference speed. EDB-Net achieves this by conducting an in-depth analysis and modeling of spatial-spectral features through two independent pipelines: one based on convolutional neural networks and the other on Transformer, thereby leveraging the complementary strengths of both approaches. Specifically, we introduce a novel lightweight spatial-spectral Transformer that incorporates a lightweight multi-head efficient attention-aware mechanism. This design ingeniously mitigates the quadratic growth of computational complexity associated with the standard self-attention mechanism's softmax calculation via the agent tokens approach. In addition, by correlating the self-attention map with the query vector, our model accurately extracts useful information to generate an attention gate that highlights key elements of the spectral sequence. Furthermore, the gated recurrent unit is incorporated into the algorithm to enhance the learning and analytical capabilities for spectral sequence data. Experimental results demonstrate that EDB-Net maintains high classification accuracy while significantly reducing computational complexity, outperforming existing state-of-the-art methods. |
|---|---|
| ISSN: | 1939-1404 2151-1535 |