Neural Architecture Search With Spatial-Spectral Attention for Higher-Order Nonlinear Hyperspectral Unmixing.

Area of Science:

• Remote Sensing
• Signal Processing
• Machine Learning

Background:

• Hyperspectral unmixing (HU) aims to identify constituent materials and their proportions in mixed pixels.
• The extended multilinear mixing (EMLM) model accurately captures complex nonlinear mixing but requires intricate deep learning architectures.
• Manual design of deep autoencoder networks for EMLM-based HU is challenging due to varying nonlinear intensities across spectral bands.

Purpose of the Study:

• To propose an automated and adaptive EMLM-based HU model using neural architecture search (NAS) with spatial-spectral attention.
• To overcome the limitations of manual network design in deep learning approaches for EMLM-based HU.
• To enhance the accuracy and efficiency of hyperspectral unmixing for complex nonlinear scenarios.

Main Methods:

• Pioneering integration of NAS into EMLM-based HU for adaptive modeling of endmember, abundance, and transition probability relationships.
• Design of a spectral-spatial attention-guided search space with multiscale convolutional operations and a NAS acceleration strategy.
• Formulation of a hybrid loss function combining linear reconstruction loss and multilinear reconstruction spectral angle distance (SAD).

Main Results:

• The proposed NAS-based EMLM HU model demonstrates significant performance advantages over existing methods.
• Adaptive modeling effectively captures intricate nonlinear mixing relationships across diverse spectral bands.
• The hybrid loss function improves convergence speed and endmember accuracy.

Conclusions:

• The developed method offers a powerful and automated solution for hyperspectral unmixing with complex nonlinear mixing.
• This approach has the potential to significantly advance the field of hyperspectral remote sensing and data analysis.
• The study provides an open-source implementation for reproducibility and further research.