GSRCA: a grad-CAM-based method for spectral region and contribution quantification in interpretability analysis.

Area of Science:

• Spectroscopy
• Chemometrics
• Machine Learning

Background:

• Standard Grad-CAM in spectral analysis suffers from fragmented gradients and limited physical interpretability.
• Existing methods lack the ability to aggregate gradient points into physically meaningful continuous spectral regions.

Purpose of the Study:

• To develop a novel method, Grad-CAM-based Spectral Region and Contribution Analysis (GSRCA), for physically interpretable spectral classification.
• To improve the aggregation of fragmented gradients and quantify the contribution of spectral regions.

Main Methods:

• GSRCA utilizes a CNN framework and Grad-CAM to compute class-specific gradients.
• An adaptive region segmentation strategy aggregates discrete gradient responses into continuous spectral regions.
• Gradient integration quantifies the relative contribution of each region to the classification result.

Main Results:

• GSRCA identified key spectral regions consistent with known molecular vibrational modes in a Raman spectroscopy dataset of illicit drugs.
• Region deletion/insertion tests validated the contribution quantification scheme.
• GSRCA demonstrated superior regional integrity, physical interpretability, and computational efficiency compared to Grad-CAM, LIME, and SHAP-based methods.

Conclusions:

• GSRCA offers a physically interpretable and versatile explanation method for spectral classification models.
• The method shows strong generalization capabilities across different architectures (CNNs, Transformers) and data types (hyperspectral imaging).
• GSRCA requires no architecture-specific parameter tuning, enhancing its applicability.