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

To address the fragmented point-wise gradients and limited physical interpretability of standard Grad-CAM in spectral analysis, this study proposes GSRCA (Grad-CAM-based Spectral Region and Contribution Analysis). Its core innovation lies in an adaptive region segmentation strategy that aggregates fragmented gradient points into continuous intervals corresponding to molecular vibrational modes. Based on a CNN framework, GSRCA first computes class-specific gradients via Grad-CAM, then employs a weight-distribution-based adaptive region segmentation strategy to aggregate discrete gradient responses into physically meaningful continuous spectral regions. Finally, the relative contribution of each region to the classification result is quantified through gradient integration and expressed as a percentage.On a Raman spectroscopy dataset containing six categories of illicit drugs, the key regions identified by GSRCA are generally consistent with characteristic vibrational bands reported in the literature, and region deletion/insertion tests further validate the effectiveness of the proposed contribution quantification scheme. Compared with conventional Grad-CAM, LIME, and SHAP-based methods such as GRASS, GSRCA shows better performance in regional integrity, physical interpretability, and computational efficiency. Experiments also show that the method requires no architecture-specific parameter tuning and can be transferred to dilated CNNs, Transformers, and remote sensing hyperspectral data, demonstrating model compatibility and generalization capability. GSRCA provides a physically interpretable and versatile explanation method for spectral classification models.