PCA-CAM: A Principal Component Analysis-Based Shell Method for Class Activation Mapping

Class Activation Mapping (CAM) methods are widely used for interpreting convolutional neural networks (CNNs) by highlighting regions in an input image that influence model predictions. Traditional CAM methods, such as Grad-CAM and Grad-CAM++, often rely on activation maps from a single layer, typically the last convolutional layer, which can overlook critical spatial information from earlier layers. This study introduces PCA-CAM, a novel method that aggregates CAM outputs across all layers using Principal Component Analysis (PCA) to produce a unified and robust explanation map. By leveraging PCA, the method integrates contributions from shallow and deep layers, capturing multi-scale features and improving the interpretability of saliency maps. Experiments on the EuroSAT dataset demonstrate that PCA-CAM consistently outperforms traditional CAM methods in generating high-quality, reliable saliency maps across diverse scenarios, while maintaining compatibility with existing architectures.