Enhancing explainability in hyperspectral target identification using PEEK-based feature attribution

Traditional deep learning models for hyperspectral imaging classification, while effective, often function as blackbox predictors with limited interpretability. This lack of transparency makes it challenging to understand how spectral features influence predictions. Traditional feature attribution methods, such as integrated gradients, provide insight but can be sensitive to baseline selection. In this work, we introduce PEEK-based explainability, an entropy-driven approach that analyzes feature importance at intermediate convolutional layers. By computing PEEK maps on spectral representations within hyperspectral imagery and projecting them onto the original spectra, we highlight the most influential spectral regions in a model's decision-making process. Our results demonstrate that PEEK visualizations provide a more intuitive and stable alternative to gradient-based attribution methods, particularly in its ability to probe intermediate layers within deep networks. This approach enhances model transparency and strengthens the connection between learned features and physical material properties, paving the way for more explainable AI in remote sensing applications.