Deep Learning-Enhanced Laser-Induced Breakdown Spectroscopy for Rapid In-Situ Analysis of Martian Surface and Atmospheric Constituents

The integration of laser-induced breakdown spectroscopy (LIBS) with deep learning algorithms presents great potential for autonomous planetary exploration. This study utilized data from the Zhurong rover to achieve spectral assignment for Martian sands, soil, and rocks. The Titanium alloy Ti-6Al-4V (TC4) based calibration strategy successfully achieves the spectral assignment of Martian atmospheric and contrast Martian and Earth spectra, distinguishing oxygen/carbon contributions. To overcome limited extraterrestrial data, we introduced the Principal Component Analysis-based Augmentation (PCA-Aug) for spectral data diversity expansion while preserving the intrinsic feature of the data. Two innovative architectures, the Lightweight Spectral Convolutional Neural Network (LS-CNN) and the Grid-Form Implicit Graph Network (GIGN), have demonstrated superior performance compared to conventional Convolutional Neural Networks (CNNs), with LS-CNN achieving a mean accuracy of 96.30% and GIGN achieving 92.59% across 10 independent training runs. These values represent consistent performance improvements of 18.4% (LS-CNN) and 13.6% (GIGN) over traditional CNNs. This study establishes LIBS as a dual-capability technique for simultaneous Martian material classification and atmospheric composition analysis, providing a critical methodology to enhance the scientific return of current and near-term Mars exploration missions.