Enhancing Transparency of Neural Networks for Super-Resolution in Remote Sensing Using Local Attribution Maps

As deep learning models advance, their use in Super Resolution (SR) tasks has become pivotal for enhancing remote sensing low-resolution (LR) satellite images. However, the decision-making processes within these neural networks remain opaque, especially in remote sensing applications where model transparency is critical. This thesis focuses on applying Explainable Artificial Intelligence (XAI) techniques, particularly Local Attribution Maps (LAMs), to analyze and interpret the internal behavior of both general-purpose and remote sensing specific SR neural networks. General-purpose models like SRGAN, EDSR, ESRT, and HAT, although highly effective in SR tasks, were originally designed for broader image enhancement challenges. In contrast, HSENet and MEN are tailored for the unique complexities of remote sensing, such as varied textures and intricate scene features. By leveraging LAMs, this research highlights how different networks prioritize and process features such as edges, textures, and high-frequency details to generate super-resolved outputs. This attribution analysis provides a detailed understanding of which input regions contribute most to the resolution enhancement, offering new insights into model behavior and performance. The comparative study between general-purpose and remote sensing-specific networks outlines the strengths and limitations of each model in handling the diverse data present in remote sensing imagery. The findings contribute to the growing field of XAI by offering transparency into SR model operations, ensuring their trustworthy application in sensitive areas like, e.g., environmental monitoring, urban planning, and reconnaissance.