# A multi-distance laser-induced breakdown spectroscopy data classification method based on deep convolutional neural network and spectral sample weight optimization

## TL;DR

This paper introduces a new method to improve the accuracy of laser-induced breakdown spectroscopy data analysis when detection distances vary, using a deep neural network with optimized sample weights.

## Contribution

A novel spectral sample weight optimization strategy is introduced to enhance CNN training for multi-distance LIBS data classification.

## Key findings

- The new weighting strategy improved testing accuracy by 8.45 percentage points compared to the original model.
- Precision, recall, and F1-score were increased by 6.4, 7.0, and 8.2 percentage points on average.
- Training time per epoch remained nearly identical to the original equal-weight scheme.

## Abstract

Laser-induced breakdown spectroscopy (LIBS) is a stand-off chemical analysis technique. In scenarios where the LIBS detection distance varies (e.g. Mars exploration), the distance effect poses a significant challenge to data analysis. In our prior work, a deep convolutional neural network (CNN) model was developed to directly process LIBS multi-distance spectra, achieving high classification accuracy even without performing conventional “distance correction”. The present study proposes a spectral sample weight optimization strategy to further improve the CNN model training process. Unlike the default equal-weight scheme, the new strategy tailors a specific weight value for every training spectral sample. On an eight-distance LIBS dataset acquired by the MarSCoDe duplicate instrument, the CNN model with the new weighting strategy can achieve a maximum testing accuracy of 92.06%, representing an improvement of 8.45 percentage points over our original CNN model. Besides accuracy, three other supplementary metrics also demonstrate the superiority of the new strategy: the precision, recall and F1-score can be averagely increased by 6.4, 7.0 and 8.2 percentage points, respectively. Moreover, the training time per epoch of the weight optimization strategy is almost identical to that of the original equal-weight scheme. These results indicate that the proposed methodology has great application potential in planetary exploration, and other LIBS-adopted scenarios involving varying detection distances.