Dynamical Hyperspectral Unmixing With Variational Recurrent Neural Networks.

Area of Science:

• Remote Sensing
• Signal Processing
• Machine Learning

Background:

• Multitemporal hyperspectral unmixing (MTHU) analyzes hyperspectral image sequences to track material (endmember) and abundance evolution.
• Existing unsupervised MTHU frameworks struggle to adequately model spatial and temporal endmember variability.

Purpose of the Study:

• To develop an unsupervised MTHU algorithm capable of effectively modeling spatial and temporal endmember variability.
• To address the limitations of current unsupervised MTHU approaches.

Main Methods:

• Proposed an unsupervised MTHU algorithm utilizing variational recurrent neural networks.
• Developed a stochastic model for endmember/abundance dynamics and mixing processes.
• Introduced a low-dimensional parametrization for spatial-temporal endmember variability.
• Formulated the problem as a Bayesian inference task solved via deep variational inference with recurrent neural networks.

Main Results:

• The proposed deep variational inference approach effectively estimates abundances and endmembers.
• The method significantly reduces the number of variables needed for endmember variability estimation.
• Experimental results demonstrate superior performance compared to state-of-the-art MTHU algorithms.

Conclusions:

• The developed variational recurrent neural network-based MTHU algorithm offers a robust solution for analyzing hyperspectral image sequences.
• The approach successfully captures complex spatial and temporal dynamics in endmember properties.
• This work advances unsupervised MTHU by providing a more comprehensive model for endmember variability.