Exploiting Digital Surface Models for Inferring Super-Resolution for Remotely Sensed Images

Despite the plethora of successful Super-Resolution Reconstruction (SRR) models applied to natural images, their application remote sensing imagery tends produce poor results. Remote is often more complicated than images and has its peculiarities such as being lower resolution, it contains noise, depicting large textured surfaces. As a result, applying non-specialized SRR like Enhanced Super Resolution Generative Adversarial Network (ESRGAN) on results in artifacts reconstructions. To address these problems, we propose novel strategy for enabling an model output realistic images: instead relying feature-space similarities perceptual loss, considers pixel-level information inferred from normalized Digital Surface Model (nDSM) image. This allows better-informed updates during training which sources task (elevation map inference) that closely related sensing. Nonetheless, nDSM auxiliary not required production i.e., infers super-resolution image without additional data. We assess our two remotely sensed datasets different spatial resolutions also contain DSMs DFC2018 dataset containing national LiDAR fly-by Luxembourg. compare with ESRGAN show achieves better performance does introduce any In particular, high-resolution are almost indistinguishable ground truth images.