Ensemble deep kernel PLS regression method for full-spectrum multi-component LIBS quantitative analysis of Mars surface oxides

Abstract

Accurately quantitative analysis of Martian surface oxides is crucial for reconstructing its geological evolution history and habitability assessment. Laser-Induced Breakdown Spectroscopy (LIBS) has demonstrated remarkable capability for quantitative compositional analysis of complex samples, such as Martian rocks and soils. However, traditional multivariate regression methods are difficult to capture the complex nonlinear relationships in LIBS data, while deep learning methods frequently suffer from high data dependency and low interpretability. To overcome this issue, this paper proposes a novel transparent and lightweight deep partial least square chemometric method inspired by deep learning, referred to as Ensemble Deep Kernel Partial Least Squares (EDKPLS). The model consists of multiple cascaded partial least squares modules with a radial basis function kernel inserted to capture nonlinear couplings between spectral lines. A latent variable ensemble strategy is utilized to alleviate the computational burden of latent variable selection and enhance predictive stability and generalization. To further improve the overall performance of the model, a multi-branch modeling approach is adopted, along with a dynamically weighted score function for parameter optimization. The proposed EDKPLS method is evaluated on the ChemCam LIBS dataset to validate its regression performance. The results across multiple analyzed components demonstrate that EDKPLS achieves lower RMSE and MAE values and higher R² compared to several representative baseline methods. Its interpretability is enabled by leveraging a back-projection method to trace the contribution of the original wavelengths to the prediction. This work presents a novel lightweight and interpretable deep modeling framework, which serves as an effective alternative chemometric tool for the analysis of Mars surface oxides using LIBS data.