SS-RAM: Spatial-Spectral Attribution Interpretability for Hyperspectral Imagery

This work proposes a novel unified explainability framework 'spatial-spectral relevance activation mapping' (SS-RAM), designed for hyperspectral images (HSIs) super-resolution (SR) models. SS-RAM integrates three attribution mechanisms: relevance conservation via layerwise relevance propagation (LRP), spatial localization through higher-order gradient-based activation mapping, and spectral fidelity via bandwise contribution analysis. These attributions are adaptively fused using a band-specific learnable coefficient, optimized through a multiobjective loss that jointly enforces fidelity, conservation, and faithfulness without manual tuning of weighting parameters. This framework is implemented as a post hoc analysis on a generative adversarial network (GAN)-based HSI-SR backbone, and SS-RAM produces interpretable heatmaps that simultaneously capture where the model attends spatially and which spectral bands drive the reconstruction. The experiments on four benchmarked HSI datasets (Indian Pines, Salinas, Pavia University, and WHU-Hi-LongKou) demonstrate near-perfect relevance conservation (≈ 0.999), high faithfulness (≈ 0.99), and balanced attribution sparsity, consistently outperforming single-source explainability methods such as LRP, Grad-CAM++, and Shapley values. The proposed framework provides robust and physically meaningful insights into deep HSI generative models there by enhancing trust and interpretability in remote sensing applications.