Unlocking Interpretability: XAI Strategies for Enhanced Insight in GNN-Based Hyperspectral Image Classification

Integrating Graph Neural Networks (GNNs) into hyperspectral (HS) image classification has led to impressive gains in accuracy. However, GNN models are complex, making it difficult to see exactly how they reach their conclusions. This study looks at how eXplainable Artificial Intelligence (XAI) techniques can make GNN-based HS classification more transparent. Methods like Grad-CAM and Grad-CAM++ are applied to gain insight into the model’s decision-making process, while Saliency Maps and Integrated Gradients show the impact of individual HS bands on those choices. Testing on the popular Indian Pines HS dataset reveals that some spectral bands are particularly influential in guiding the model’s classifications. By improving model transparency, this work aims to increase trust in GNNs and support their broader use in remote sensing applications.