Shap Value-Guided Selection of Spectral Variables in Libs for Improving Soil Carbon Determination

In laser-induced breakdown spectroscopy (LIBS), identifying key emission lines for accurate elemental quantification has long posed a challenge. Traditional methods rely on experimental knowledge, atomic databases, and intricate spectral analyses. Although machine learning techniques – such as boosting algorithms and neural networks – offer efficient processing for large datasets, the complexity of these techniques often compromises interpretability. To address this issue, our study integrates the SHapley Additive exPlanations (SHAP) algorithm with gradient boosting models in order to refine the selection of spectral variables, thus enhancing our understanding of how specific emission lines contribute to the carbon (C) concentration predictions in soils. Deployed on a large dataset of 1,019 soil samples, a wrapper method with a Random Forest regressor reduced the initial spectral intensity variables from 13,748 to 1,098. The subsequent application of a LightGBM regressor calibrated via the Optuna framework yielded – for training and validation sets, respectively – an R2 of 0.98 and 0.77, and RMSE values of 1.55 and 4.54 g kg−1. The SHAP summary plot showed that C emission lines influenced the model’s predictions positively, as anticipated, whereas silicon (Si) emission lines produced a negative impact, suggesting a sandy soils with lower C concentration. Not only do our findings validate the efficacy of the SHAP method in enhancing LIBS-based soil C quantification, but they also offer a sophisticated framework for decoding the complex interplay between emission lines and target elemental concentrations.