Enhancing explainability in Raman spectroscopy classification with SHAP and spectral segmentation.

Area of Science:

• Spectroscopy
• Chemometrics
• Machine Learning

Background:

• Interpreting spectral classification models is challenging.
• Traditional methods often focus on individual data points, missing regional spectral information.
• Need for methods that provide physically meaningful explanations for spectral classification decisions.

Purpose of the Study:

• Introduce GRASS (Gradient-Region Analyzed Spectral SHAP), a new method to enhance spectral classification interpretability.
• Improve transparency and physical interpretability of Raman spectroscopy classification.
• Validate the physical rationale of identified spectral regions against molecular vibration modes.

Main Methods:

• Integrate gradient-region analysis with spectral SHAP value quantification.
• Automatically segment high-dimensional spectral data into continuous regions using backpropagation gradients.
• Apply SHAP to quantify the contribution of each identified spectral region to classification.

Main Results:

• GRASS identified key spectral regions that correspond to molecular vibration modes.
• GRASS demonstrated better alignment with synergistic spectral region interactions than Grad-CAM and LIME.
• GRASS showed robustness across different data normalization, network architectures, and hyperspectral datasets.

Conclusions:

• GRASS provides enhanced interpretability for spectral classification models.
• The method aligns with physical principles, offering physically meaningful explanations.
• GRASS is suitable for high-reliability fields and can be extended to other spectral types.