.png){kind=icon}
.png){kind=icon}
Return
Hyperspectral unmixing algorithm based on channel multi-scale dual-stream autoencode
.png){kind=icon}
PRE
AI
DOI：10.1080/01431161.2025.2463699.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Autoencoders (AEs) have made significant progress for hyperspectral unmixing. However, traditional AE models exhibit some deficiencies in capturing and integrating global spatial and spectral information for hyperspectral unmixing. To overcome these limitations, this paper proposes a Swin Transformer-based Dual-Stream Self-Attention Autoencoder Network (SDSAN), aiming to enhance the capabilities of extracting spatial and spectral features. First, SDSAN adopts a dual-stream architecture, dividing the feature extraction process into spatial stream and spectral stream. The two streams can independently analyse the data features from spatial and spectral perspectives, respectively. This structure can enhance the efficiency of feature extraction. Second, SDSAN integrates two key modules: the Spatial Feature Extraction Module (SFEM) and the Multi-Scale Spectral Attention Block (MSSAB). The SFEM uses the Swin Transformer and grouped convolutions to capture multiscale spatial features effectively. Meanwhile, the MSSAB utilizes dilated convolutions and a spectral self-attention mechanism to extract spectral features across multiple scales. This dual-stream structure significantly improve the network's ability to unmix complex hyperspectral imagery. Finally, extensive experiments have conducted on synthetic and real hyperspectral datasets. We compare SDSAN with five benchmark algorithms, and the experimental results demonstrate the good performance of SDSAN.
Keywords:
Hyperspectral unmixing
dual-stream autoencoder
Swin Transformer
grouped convolution
Journal
IF:
2.6
Papers:
1.2W
Citations:
2.7W