Hyperspectral Unmixing Network Inspired by Unfolding an Optimization Problem

Chao ZhouarXiv 2020 · arXiv

TL;DR

This paper introduces novel neural network architectures for hyperspectral unmixing by "unfolding" the Alternating Direction Method of Multipliers (ADMM) optimization algorithm. This approach bridges the gap between interpretable model-based methods and powerful but opaque deep learning models, achieving faster convergence and competitive performance with less training data.

What Problem Does This Paper Address?

Hyperspectral Unmixing (HSU) aims to identify the constituent materials (endmembers) and their proportions (abundances) within each pixel of a hyperspectral image (HSI). Traditional HSU methods often rely on iterative optimization algorithms, which are computationally expensive. While deep learning offers faster solutions, these "black box" models often lack interpretability and require large labeled datasets, which are difficult to obtain for HSI. This paper addresses the need for HSI unmixing methods that are both interpretable and efficient, combining the strengths of model-based optimization and data-driven learning.

Methods

The core innovation lies in "unfolding" the ADMM algorithm, a popular optimization technique for HSI unmixing, into a neural network architecture. The authors break down each iterative step of ADMM into distinct network layers: an X-update layer, a Z-update layer, and a D-update layer. These layers are designed to learn parameters that were previously fixed in the ADMM algorithm.

Two main network architectures are proposed:

  • U-ADMM-AENet: For abundance estimation. It comes in two versions:
  • U-ADMM-AENet-I: Uses shared layers across iteration blocks, similar to Recurrent Neural Networks (RNNs), promoting interpretability.
  • U-ADMM-AENet-II: Uses untied layers for each block, offering greater flexibility and approximation capability, functioning more like a feed-forward network.
  • U-ADMM-BUNet: For blind unmixing (where endmembers are also unknown). This architecture adds a linear decoder layer to the abundance estimation network.

Key methodological aspects include:

  • Interpretability: The unfolded layers have direct interpretations related to ADMM steps, allowing for better understanding and initialization.
  • Initialization Strategies: Novel strategies are proposed to leverage prior knowledge about endmembers and optimization parameters, providing a "warm start" for the networks.
  • Custom Loss Function: Beyond Mean Squared Error (MSE), the authors introduce Abundance Angle Distance (AAD) and Abundance Information Divergence (AID) to better capture dissimilarity in abundance estimates.
  • Residual Connections: The unfolded ADMM structure naturally incorporates residual connections, a known beneficial feature in deep learning for image processing.

Datasets

The paper uses both synthetic and real-world hyperspectral datasets for evaluation:

  • Synthetic Data: Generated by creating endmember signatures and then convolving abundance maps with Gaussian filters, followed by the addition of additive white Gaussian noise (AWGN) at various Signal-to-Noise Ratios (SNRs).
  • Real Datasets:
    • Jasper Ridge: A widely used HSI dataset.
  • Urban: Another common HSI dataset.

Key Insights and Ideas

  • Unfolding ADMM for Neural Networks: The paper demonstrates that a powerful optimization algorithm like ADMM can be directly translated into a learnable neural network architecture, offering a principled way to design deep learning models for inverse problems.
  • Bridging the Gap: This approach successfully combines the interpretability and prior knowledge integration of model-based methods with the learning capacity and speed of deep learning.
  • Interpretability Drives Initialization: The direct link to ADMM steps allows for informed initialization strategies, which significantly improve training efficiency and performance, especially with limited data.
  • Residual Structure is Natural: The ADMM unfolding naturally leads to a network structure with residual connections, which are known to enhance deep learning model performance.
  • Beyond MSE: Using a combination of MSE, AAD, and AID as loss functions provides a more comprehensive measure of abundance estimation quality, leading to better results.

Key Results

  • Faster Convergence: The proposed U-ADMM-AENet models achieve convergence with significantly fewer training epochs (e.g., 100 epochs) compared to other learning-based methods like MNN-AE (e.g., 300 epochs).
  • Improved Performance with Less Data: The networks demonstrate competitive and often superior performance even when trained with very small datasets, a crucial advantage for HSI applications where labeled data is scarce.
  • Robustness to Noise: The proposed methods show better performance and stability under noisy conditions compared to some existing algorithms.
  • Competitive Performance on Real Data: On the Jasper Ridge and Urban datasets, the U-ADMM-AENet models achieve state-of-the-art or near state-of-the-art results in abundance estimation, outperforming traditional and other learning-based methods.
  • Effective Blind Unmixing: The U-ADMM-BUNet models also show promising results in blind unmixing, demonstrating good performance with less training data and faster convergence.

Limitations and Future Work

  • Shared vs. Untied Layers: While U-ADMM-AENet-I (shared layers) is more interpretable, it can be less stable with increasing iteration blocks. U-ADMM-AENet-II (untied layers) offers more flexibility but might be less interpretable. The paper notes that for U-ADMM-AENet-I, better performance can be achieved with just one iteration block.
  • Endmember Signature Estimation in Blind Unmixing: For the Jasper Ridge dataset, the U-ADMM-BUNet-II does not outperform other methods in terms of individual endmember signatures, though it excels in average SAD performance.
  • Future Work: The paper suggests exploring more complex mixture models beyond the linear model and investigating the impact of different initialization strategies and loss functions further. The authors also hint at the potential for further optimization of the network architecture and training procedures.

Summary updated

Abstract

The hyperspectral image (HSI) unmixing task is essentially an inverse problem, which is commonly solved by optimization algorithms under a predefined (non-)linear mixture model. Although these optimization algorithms show impressive performance, they are very computational demanding as they often rely on an iterative updating scheme. Recently, the rise of neural networks has inspired lots of learning based algorithms in unmixing literature. However, most of them lack of interpretability and require a large training dataset. One natural question then arises: can one leverage the model based algorithm and learning based algorithm to achieve interpretable and fast algorithm for HSI unmixing problem? In this paper, we propose two novel network architectures, named U-ADMM-AENet and U-ADMM-BUNet, for abundance estimation and blind unmixing respectively, by combining the conventional optimization-model based unmixing method and the rising learning based unmixing method. We first consider a linear mixture model with sparsity constraint, then we unfold Alternating Direction Method of Multipliers (ADMM) algorithm to construct the unmixing network structures. We also show that the unfolded structures can find corresponding interpretations in machine learning literature, which further demonstrates the effectiveness of proposed methods. Benefit from the interpretation, the proposed networks can be initialized by incorporating prior information about the HSI data. Different from traditional unfolding networks, we propose a new training strategy for proposed networks to better fit in the HSI applications. Extensive experiments show that the proposed methods can achieve much faster convergence and competitive performance even with very small size of training data, when compared with state-of-art algorithms.
BibTeX
@misc{zhou2020hyperspectral,
  author = {Chao Zhou},
  title = {{Hyperspectral Unmixing Network Inspired by Unfolding an Optimization Problem}},
  year = {2020},
  eprint = {2005.10856},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2005.10856},
}
Figure