A multi-distance laser-induced breakdown spectroscopy data classification method based on deep convolutional neural...

Area of Science:

• Planetary Science
• Analytical Chemistry
• Spectroscopy

Background:

• Laser-Induced Breakdown Spectroscopy (LIBS) is a vital stand-off chemical analysis technique.
• Varying detection distances in applications like Mars exploration present significant challenges for LIBS data analysis.
• Previous deep convolutional neural network (CNN) models processed multi-distance LIBS spectra effectively without distance correction.

Purpose of the Study:

• To introduce and evaluate a spectral sample weight optimization strategy for enhancing CNN model training in LIBS.
• To improve the classification accuracy and performance metrics of LIBS analysis across varying distances.
• To assess the computational efficiency of the proposed weighting strategy.

Main Methods:

• Developed a spectral sample weight optimization strategy for CNN model training.
• Applied the strategy to an eight-distance LIBS dataset from the MarSCoDe duplicate instrument.
• Compared the performance of the optimized CNN model against the original model using accuracy, precision, recall, and F1-score.

Main Results:

• The CNN model with the spectral sample weight optimization achieved a maximum testing accuracy of 92.06%, an 8.45 percentage point improvement.
• Precision, recall, and F1-score saw average increases of 6.4, 7.0, and 8.2 percentage points, respectively.
• The training time per epoch remained comparable to the original equal-weight scheme.

Conclusions:

• The proposed spectral sample weight optimization strategy significantly enhances LIBS analysis accuracy and performance.
• This methodology offers a promising solution for LIBS applications with varying detection distances, particularly in planetary exploration.
• The strategy provides superior results without increasing computational training time, demonstrating its practical applicability.