Spatial-Channel Multiscale Transformer Network for Hyperspectral Unmixing.

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Deep learning (DL) shows promise in hyperspectral unmixing (HU).
• Convolutional neural networks (CNNs) capture local spatial data but struggle with long-range dependencies.
• Transformers excel at global context but often lack unified spatial-channel and multiscale modeling.

Purpose of the Study:

• To propose a novel Spatial-Channel Multiscale Transformer Network (SCMT-Net) for hyperspectral unmixing.
• To address limitations in existing transformer-based HU methods regarding multiscale spatial and channel-wise dependency modeling.
• To enhance feature representation and contextual understanding in complex hyperspectral scenes.

Main Methods:

• A compact feature projection (CFP) module for initial feature extraction.
• Sequential application of spatial multiscale transformer (SMT) and channel multiscale transformer (CMT) for modeling spatial and spectral dependencies.
• Integration of multiscale multi-head self-attention (MMSA) and an efficient feed-forward network (E-FFN) for enhanced feature fusion and information flow.

Main Results:

• SCMT-Net demonstrated superior performance in both abundance estimation and endmember extraction across multiple real and synthetic hyperspectral datasets.
• The proposed network effectively captures multiscale spatial and channel-wise dependencies.
• Experiments on Samson, Jasper, and Apex datasets confirmed the robustness and accuracy of SCMT-Net.

Conclusions:

• SCMT-Net offers a significant advancement in hyperspectral unmixing by effectively integrating multiscale spatial and channel-wise contextual information.
• The model achieves a favorable balance between accuracy and computational efficiency.
• SCMT-Net represents a robust and accurate solution for complex hyperspectral data analysis.