Adversarial Autoencoder Network for Hyperspectral Unmixing.

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Spectral unmixing (SU) is crucial for hyperspectral image analysis, extracting subpixel features and abundances.
• Autoencoders (AEs) are used for unsupervised SU but struggle with prior properties, noise, and initialization.
• Existing methods have limitations in exploiting spatial correlations and prior distributions for accurate unmixing.

Purpose of the Study:

• To propose a novel unsupervised spectral unmixing technique using an adversarial autoencoder network (AAENet).
• To enhance the accuracy and robustness of hyperspectral unmixing by integrating spatial information and abundance priors.
• To address limitations of traditional AEs in exploiting prior properties and handling noise.

Main Methods:

• Developed an adversarial autoencoder network (AAENet) for unsupervised spectral unmixing.
• Modeled abundance to follow a prior distribution by assuming homogeneous regions share statistical properties.
• Employed adversarial training to transfer spatial information and match abundance posterior with a prior distribution.

Main Results:

• The proposed AAENet demonstrated more accurate and interpretable unmixing performance.
• AAENet significantly enhanced model performance and robustness compared to traditional AE methods.
• Experiments on simulated and real hyperspectral data showed superior results over state-of-the-art methods.

Conclusions:

• The AAENet effectively leverages spatial correlations and abundance priors for improved unsupervised spectral unmixing.
• Adversarial training enhances the ability to incorporate prior knowledge, leading to better unmixing outcomes.
• AAENet offers a robust and accurate solution for hyperspectral image analysis challenges.