# Target-enhanced double-pulse LIBS coupled with feature-fused CNN for mechanistic and interpretable coffee origin authentication

## TL;DR

A new method combining laser spectroscopy and machine learning accurately identifies coffee origins, outperforming traditional techniques.

## Contribution

A novel framework integrating potassium-assisted DP-LIBS with a Feature-Fused CNN for interpretable and accurate coffee origin authentication.

## Key findings

- The Feature-Fused CNN achieved 99.00% classification accuracy for coffee origin.
- The model retained >94% accuracy under 30 dB SNR, showing strong anti-interference performance.
- SHAP and 1D Grad-CAM++ revealed the model's reliance on synergistic trace elements like Fe, Cr, Cu, and K.

## Abstract

Coffee geographical origin authentication is critical for mitigating economically motivated adulteration, yet rapid trace-element analysis in complex organic matrices remains a significant challenge. This study establishes a novel synergistic framework integrating Potassium-assisted orthogonal Double-Pulse Laser-Induced Breakdown Spectroscopy (DP-LIBS) with a Feature-Fused CNN for precise coffee traceability. A high-purity KHCO₃ solid target was employed to facilitate plasma cross-coupling and secondary energy injection, significantly enhancing signal sensitivity. Surmounting the inherent bottlenecks of manual feature engineering, a Feature-Fused CNN architecture was constructed by concatenating normalized spectral data with statistical descriptors, enabling the autonomous extraction of hierarchical spatial-spectral patterns. The proposed model achieved a superior classification accuracy and F1-score of 99.00%, significantly outperforming traditional algorithms including XGBoost (95.75%), PLS-DA (92.50%), RF (86.50%), and KNN (75.75%). Robustness evaluation demonstrated that the CNN maintained high precision (>94%) even under severe noise interference (30 dB SNR). Furthermore, a dual-interpretability strategy was implemented to elucidate the decision logic: SHAP analysis was utilized to quantify feature contributions for traditional machine learning models, identifying key markers such as Fe, Cr, and Na; meanwhile, 1D Grad-CAM++ was applied to the Feature-Fused CNN to visualize wavelength-specific activation weights. The results reveal that the CNN's superior performance stems from recognizing the synergistic covariance of trace elements (Fe, Cr, Cu, and K) rather than isolated spectral peaks, providing a robust and mechanically interpretable strategy for food provenance verification.

•Potassium-assisted orthogonal DP-LIBS coupled with Feature-Fused CNN achieves 99.00% coffee origin classification accuracy, outperforming traditional algorithms.•KHCO₃ target-enhanced orthogonal DP-LIBS realizes plasma cross-coupling, markedly boosting trace element signal sensitivity in complex coffee matrices.•The Feature-Fused CNN fuses spectral and statistical features, retaining >94% accuracy under 30 dB SNR with exceptional anti-interference performance.•Dual interpretability via SHAP and 1D Grad-CAM++ reveals model’s reliance on Fe/Cr/Cu/K synergistic covariance, addressing deep learning “black box” issues.•This work establishes a generalizable paradigm for trace element analysis in complex food matrices, with broad industrial application potential.

Potassium-assisted orthogonal DP-LIBS coupled with Feature-Fused CNN achieves 99.00% coffee origin classification accuracy, outperforming traditional algorithms.

KHCO₃ target-enhanced orthogonal DP-LIBS realizes plasma cross-coupling, markedly boosting trace element signal sensitivity in complex coffee matrices.

The Feature-Fused CNN fuses spectral and statistical features, retaining >94% accuracy under 30 dB SNR with exceptional anti-interference performance.

Dual interpretability via SHAP and 1D Grad-CAM++ reveals model’s reliance on Fe/Cr/Cu/K synergistic covariance, addressing deep learning “black box” issues.

This work establishes a generalizable paradigm for trace element analysis in complex food matrices, with broad industrial application potential.