Publication detail

Classification of spectroscopic data - challenges, benchmarking and limitations

VRÁBEL, J. KÉPEŠ, E. POŘÍZKA, P. KAISER, J.

Original Title

Classification of spectroscopic data - challenges, benchmarking and limitations

Type

abstract

Language

English

Original Abstract

In modern spectroscopy, we are often dealing with large and highly complex datasets. As an example, in Laser-Induced Breakdown Spectroscopy (LIBS), measurements with a 1 kHz repetition rate were reported. Such a measurement often results in huge, high-dimensional data that are impossible to explore and analyze by a hand. Even more, many common methods (Principal Component Analysis + classifier, Support Vector Machines, etc.) may become insufficient and new strategies are required. Classification of large spectroscopic data is a challenging task due to the nature of spectra. Modern machine learning (ML) techniques based on artificial neural networks (ANN) are opening new possibilities, but often there is a lack of understanding in the decision processes (for classification). In this work, we extensively study modern approaches to classification with a focus on the explainability of decision factors. Innovative models with the incorporation of physics (or spectra modeling) are discussed. Besides mentioned, we focus on the extendability of the approaches beyond the LIBS method.

Keywords

spectroscopic data, machine learning, benchmarking, classification

Authors

VRÁBEL, J.; KÉPEŠ, E.; POŘÍZKA, P.; KAISER, J.

Released

8. 7. 2020

BibTex

@misc{BUT165756,
  author="Jakub {Vrábel} and Erik {Képeš} and Pavel {Pořízka} and Jozef {Kaiser}",
  title="Classification of spectroscopic data - challenges, benchmarking and limitations",
  year="2020",
  note="abstract"
}