Coherent Spectral Feature Extraction Using Symmetric Autoencoders

TL;DR

This paper introduces SymAE, a symmetric autoencoder that disentangles class-invariant spectral features from nuisance variability in hyperspectral images, improving land-cover classification especially in disjoint training-test scenarios.

Contribution

The paper presents a novel data-driven autoencoder architecture that disentangles coherent spectral features without handcrafted models or priors, enhancing hyperspectral classification.

Findings

SymAE achieves state-of-the-art classification accuracy.

Virtual spectra generation offers interpretability insights.

Improves performance in disjoint training-test scenarios.

Abstract

Hyperspectral data acquired through remote sensing are invaluable for environmental and resource studies. While rich in spectral information, various complexities such as environmental conditions, material properties, and sensor characteristics can cause significant variability even among pixels belonging to the same material class. This variability poses nuisance for accurate land-cover classification and analysis. Focusing on the spectral domain, we utilize an autoencoder architecture called the symmetric autoencoder (SymAE), which leverages permutation invariant representation and stochastic regularization in tandem to disentangle class-invariant 'coherent' features from variability-causing 'nuisance' features on a pixel-by-pixel basis. This disentanglement is achieved through a purely data-driven process, without the need for hand-crafted modeling, noise distribution priors, or reference 'clean signals'. Additionally, SymAE can generate virtual spectra through manipulations in latent space. Using AVIRIS instrument data, we demonstrate these virtual spectra, offering insights on the disentanglement. Extensive experiments across six benchmark hyperspectral datasets show that coherent features extracted by SymAE can be used to achieve state-of-the-art pixel-based classification. Furthermore, we leverage these coherent features to enhance the performance of some leading spectral-spatial HSI classification methods. Our approach especially shows improvement in scenarios where training and test sets are disjoint, a common challenge in real-world applications where existing methods often struggle to maintain relatively high performance.