HYPERION: Hyperspectral Penetrating-type Ellipsoidal Reconstruction for Terahertz Blind Source Separation

Chia-Hsiang Lin, Yi-Chun Hung, Fengyu Wang, Shang-Hua YangarXiv 2021 · DOI · arXiv

Abstract

Terahertz (THz) technology has been a great candidate for applications, including pharmaceutic analysis, chemical identification, and remote sensing and imaging due to its non-invasive and non-destructive properties. Among those applications, penetrating-type hyperspectral THz signals, which provide crucial material information, normally involve a noisy, complex mixture system. Additionally, the measured THz signals could be ill-conditioned due to the overlap of the material absorption peak in the measured bands. To address those issues, we consider penetrating-type signal mixtures and aim to develop a blind hyperspectral unmixing (HU) method without requiring any information from a prebuilt database. The proposed HYperspectral Penetrating-type Ellipsoidal ReconstructION (HYPERION) algorithm is unsupervised, not relying on collecting extensive data or sophisticated model training. Instead, it is developed based on elegant ellipsoidal geometry under a very mild requirement on data purity, whose excellent efficacy is experimentally demonstrated.

Introduction

Terahertz technology offers a unique window into the non-destructive analysis of materials. However, interpreting the data from penetrating-type THz signals is challenging because the resulting spectra are often complex mixtures of multiple chemical signatures. Traditional blind source separation techniques, such as Independent Component Analysis (ICA) and Non-negative Matrix Factorization (NMF), frequently struggle with the statistical dependencies of chemical mixtures, sensitivity to noise, and the requirement for high-purity data points. Furthermore, THz data is often "ill-conditioned," where different chemicals exhibit overlapping absorption peaks, making them difficult to distinguish.

The HYPERION Algorithm

To overcome these limitations, we introduce HYPERION (HYperspectral Penetrating-type Ellipsoidal ReconstructION). Unlike statistical methods, HYPERION treats unmixing as a geometric problem. The core of the algorithm is the Löwner-John Ellipsoid (LJE), which is used to reconstruct the source signatures from mixed data.

  1. Standardization: Raw THz transmission data is converted into material absorption coefficients.
  2. Affine Fitting: High-dimensional data is projected into a lower-dimensional convex hull using Singular Value Decomposition (SVD) to mitigate noise.
  3. Ellipsoidal Geometry: The algorithm identifies the maximum-volume ellipsoid (LJE) that fits within the data's geometric shape.
  4. Preconditioning: By transforming this ellipsoid into a Euclidean ball, the algorithm "stretches" the data, effectively magnifying the subtle differences between ill-conditioned chemical signatures.
  5. Simplex Fitting: A regular simplex is fitted to the data, where the vertices represent the "pure" chemical signatures.

This unsupervised approach does not require a pre-existing database or extensive training, relying instead on the inherent geometric structure of the mixture data.

Experimental Validation

HYPERION was validated using two primary experimental setups involving ternary and quinary chemical mixtures. The ternary dataset included D-lactose monohydrate, D-sucrose, and L-tyrosine. The more complex quinary dataset added D-glucose and L-histidine. Notably, D-glucose and D-sucrose have highly similar THz profiles, serving as a rigorous test for the algorithm's ability to handle ill-conditioned data.

Experimental validation of HYPERION on ternary mixtures showing power spectra and unmixing performance metrics.

Key Results and Discussion

HYPERION demonstrated significant improvements over traditional methods. In the quinary test, the algorithm achieved a Spectral Angle Mapper (SAM) error of only 9.57 degrees and a Root Mean Square Error (RMSE) of 2.30 cm⁻¹, outperforming both NMF and ICA.

One of the most critical insights is the power of preconditioning. By using the LJE to normalize the data, HYPERION can distinguish between chemicals with overlapping spectra, such as glucose and sucrose, which would otherwise be indistinguishable in noisy environments. The algorithm is also remarkably efficient, converging in less than 3 seconds on standard consumer hardware.

Unmixing results for the quinary dataset and comparison of overlapping absorption profiles for D-glucose and D-sucrose.

In a real-world "Blind Mapping Test," HYPERION successfully visualized the composition of 15 hidden tablets behind copper foil with less than 20% inaccuracy, proving its potential for remote chemical identification and "secret recipe" detection.

Conclusion and Future Work

HYPERION provides a robust, unsupervised framework for THz blind source separation. While highly effective, current limitations include accuracy drops at frequencies above 1.5 THz due to decreasing signal-to-noise ratios and a dependency on known sample thickness. Future research will focus on estimating the complex refractive index simultaneously and incorporating spatial regularizers to improve imaging results by leveraging the logical consistency of neighboring pixels.

Summary updated

Abstract

Terahertz (THz) technology has been a great candidate for applications, including pharmaceutic analysis, chemical identification, and remote sensing and imaging due to its non-invasive and non-destructive properties. Among those applications, penetrating-type hyperspectral THz signals, which provide crucial material information, normally involve a noisy, complex mixture system. Additionally, the measured THz signals could be ill-conditioned due to the overlap of the material absorption peak in the measured bands. To address those issues, we consider penetrating-type signal mixtures and aim to develop a blind hyperspectral unmixing (HU) method without requiring any information from a prebuilt database. The proposed HYperspectral Penetrating-type Ellipsoidal ReconstructION (HYPERION) algorithm is unsupervised, not relying on collecting extensive data or sophisticated model training. Instead, it is developed based on elegant ellipsoidal geometry under a very mild requirement on data purity, whose excellent efficacy is experimentally demonstrated.
BibTeX
@misc{lin2021hyperion,
  author = {Chia-Hsiang Lin and Yi-Chun Hung and Fengyu Wang and Shang-Hua Yang},
  title = {{HYPERION: Hyperspectral Penetrating-type Ellipsoidal Reconstruction for Terahertz Blind Source Separation}},
  year = {2021},
  eprint = {2109.05425},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2109.05425},
}
Figure