A matrix-tolerant one-dimensional CNN for raw LIBS analysis: Architecture and soil carbon benchmark

Matrix effects, line overlap, and weak signals limit quantitative accuracy in Laser-Induced Breakdown Spectroscopy (LIBS), especially for heterogeneous samples such as soils. Conventional pipelines rely on extensive preprocessing and hand-crafted features, which can attenuate informative but low-contrast regions and reduce transferability across matrices. Here, a one-dimensional convolutional neural network (1D-CNN) is designed to learn directly from raw spectra, minimizing manual intervention while preserving full spectral structure. The driving problem addressed is whether a raw-spectra CNN can deliver externally validated, physically consistent carbon quantification in soils despite matrix variation and line interference.

A large experimental dataset comprising 1019 Brazilian soil samples was used for model development, with generalization assessed using an independent, temporally separated external validation set (n = 150). The proposed raw-spectra CNN outperformed classical machine-learning baselines and a feature-selected CNN configuration based on Minimum Redundancy Maximum Relevance (MRMR) method. On the external dataset, the model achieved an R² of 0.81, RMSE of 5.18 g kg^-1, and RPIQ of 4.65. Explicit feature selection led to a systematic reduction in external performance, suggesting that limiting the input space discarded weak but informative spectral cues. Beyond global accuracy, repeatability and normalized error analyses demonstrated stable analytical behavior across carbon concentration ranges, while an estimated limit of detection of 3.7 g kg^-1 was consistent with practical LIBS constraints. Local Interpretable Model-Agnostic Explanations (LIME) highlighted spectral regions aligned with carbon emission lines and matrix-element transitions, supporting physics-consistent signal use rather than spurious correlations. Collectively, these results show that preserving the complete spectral manifold enables the CNN to learn local/overlapping lines alongside broad contextual structure, reducing sensitivity to matrix effects and improving external transferability.

This work demonstrates that convolutional neural networks trained on raw LIBS spectra can provide a robust and analytically meaningful framework for soil carbon quantification, combining external validation, controlled repeatability, and spectroscopically plausible interpretability. By reducing dependence on preprocessing and feature engineering while preserving generalization, the proposed approach supports simpler and more transparent LIBS workflows for complex matrices. The methodology is broadly applicable and may be extended to other elements, soil properties, and LIBS-based analytical scenarios.