Activation Map-Based Visual Explanation for Remote Sensing Image Super-Resolution

Deep learning has enhanced image super-resolution (SR) performance, yet SR networks are often treated as closed boxes with limited explainability. Notably, remote sensing (RS) images possess explicit physical semantics, imposing stringent demands on the explainability of their SR methods. To address this, we propose AMVE-SR, a framework that systematically transfers activation map-based visual explanation (AMVE) from classification networks to SR networks. Distinct from classification-oriented methods (e.g., Grad-CAM) that struggle with shallow-layer visualization or computationally intensive perturbation-based approaches, AMVE-SR introduces a layer-wise recursive propagation mechanism. By exploiting interlayer correlations, it traces activation maps backward to the input, ensuring high-fidelity explanations even in shallow layers where conventional methods fail to transfer effectively. Furthermore, we compute gradient responses via textural discrepancy rather than raw pixel intensities, capturing high-frequency details (e.g., physical boundaries) essential for physical semantics. To address the lack of quantitative metrics, we propose average texture variation (ATV) by adapting the average drop concept in classification tasks. Validated by visual consistency, ATV serves as a novel fidelity metric for SR explanations. Experiments on the AID dataset using the EDSR model demonstrate that our method produces visually reliable explanation maps while achieving superior performance on the proposed ATV metric with high running efficiency. In general, this study extends AMVE from image classification to SR, thereby establishing a new paradigm to explain SR networks.